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 "packet_internal.h"
  29 #include "snow_dwt.h"
  30 #include "snow.h"
  31
  32 #include "rangecoder.h"
  33 #include "mathops.h"
  34
  35 #include "mpegvideo.h"
  36 #include "h263.h"
  37
  38 static av_cold int encode_init(AVCodecContext *avctx)
  39 {
  40     SnowContext *s = avctx->priv_data;
  41     int plane_index, ret;
  42     int i;
  43
  44     if(s->pred == DWT_97
  45        && (avctx->flags & AV_CODEC_FLAG_QSCALE)
  46        && avctx->global_quality == 0){
  47         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
  48         return AVERROR(EINVAL);
  49     }
  50
  51     s->spatial_decomposition_type= s->pred; //FIXME add decorrelator type r transform_type
  52
  53     s->mv_scale       = (avctx->flags & AV_CODEC_FLAG_QPEL) ? 2 : 4;
  54     s->block_max_depth= (avctx->flags & AV_CODEC_FLAG_4MV ) ? 1 : 0;
  55
  56     for(plane_index=0; plane_index<3; plane_index++){
  57         s->plane[plane_index].diag_mc= 1;
  58         s->plane[plane_index].htaps= 6;
  59         s->plane[plane_index].hcoeff[0]=  40;
  60         s->plane[plane_index].hcoeff[1]= -10;
  61         s->plane[plane_index].hcoeff[2]=   2;
  62         s->plane[plane_index].fast_mc= 1;
  63     }
  64
  65     if ((ret = ff_snow_common_init(avctx)) < 0) {
  66         return ret;
  67     }
  68     ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
  69
  70     ff_snow_alloc_blocks(s);
  71
  72     s->version=0;
  73
  74     s->m.avctx   = avctx;
  75     s->m.bit_rate= avctx->bit_rate;
  76     s->m.lmin    = avctx->mb_lmin;
  77     s->m.lmax    = avctx->mb_lmax;
  78     s->m.mb_num  = (avctx->width * avctx->height + 255) / 256; // For ratecontrol
  79
  80     s->m.me.temp      =
  81     s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
  82     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  83     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  84     s->m.sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
  85     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.sc.obmc_scratchpad)
  86         return AVERROR(ENOMEM);
  87
  88     ff_h263_encode_init(&s->m); //mv_penalty
  89
  90     s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
  91
  92     if(avctx->flags&AV_CODEC_FLAG_PASS1){
  93         if(!avctx->stats_out)
  94             avctx->stats_out = av_mallocz(256);
  95
  96         if (!avctx->stats_out)
  97             return AVERROR(ENOMEM);
  98     }
  99     if((avctx->flags&AV_CODEC_FLAG_PASS2) || !(avctx->flags&AV_CODEC_FLAG_QSCALE)){
 100         ret = ff_rate_control_init(&s->m);
 101         if(ret < 0)
 102             return ret;
 103     }
 104     s->pass1_rc= !(avctx->flags & (AV_CODEC_FLAG_QSCALE|AV_CODEC_FLAG_PASS2));
 105
 106     switch(avctx->pix_fmt){
 107     case AV_PIX_FMT_YUV444P:
 108 //    case AV_PIX_FMT_YUV422P:
 109     case AV_PIX_FMT_YUV420P:
 110 //    case AV_PIX_FMT_YUV411P:
 111     case AV_PIX_FMT_YUV410P:
 112         s->nb_planes = 3;
 113         s->colorspace_type= 0;
 114         break;
 115     case AV_PIX_FMT_GRAY8:
 116         s->nb_planes = 1;
 117         s->colorspace_type = 1;
 118         break;
 119 /*    case AV_PIX_FMT_RGB32:
 120         s->colorspace= 1;
 121         break;*/
 122     default:
 123         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
 124         return AVERROR_PATCHWELCOME;
 125     }
 126
 127     ret = av_pix_fmt_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift,
 128                                            &s->chroma_v_shift);
 129     if (ret) {
 130         av_log(avctx, AV_LOG_ERROR, "pixel format invalid or unknown\n");
 131         return ret;
 132     }
 133
 134     ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
 135     ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
 136
 137     s->input_picture = av_frame_alloc();
 138     if (!s->input_picture)
 139         return AVERROR(ENOMEM);
 140
 141     if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
 142         return ret;
 143
 144     if(s->motion_est == FF_ME_ITER){
 145         int size= s->b_width * s->b_height << 2*s->block_max_depth;
 146         for(i=0; i<s->max_ref_frames; i++){
 147             s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
 148             s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
 149             if (!s->ref_mvs[i] || !s->ref_scores[i])
 150                 return AVERROR(ENOMEM);
 151         }
 152     }
 153
 154     return 0;
 155 }
 156
 157 //near copy & paste from dsputil, FIXME
 158 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
 159 {
 160     int s, i, j;
 161
 162     s = 0;
 163     for (i = 0; i < h; i++) {
 164         for (j = 0; j < w; j++) {
 165             s += pix[0];
 166             pix ++;
 167         }
 168         pix += line_size - w;
 169     }
 170     return s;
 171 }
 172
 173 //near copy & paste from dsputil, FIXME
 174 static int pix_norm1(uint8_t * pix, int line_size, int w)
 175 {
 176     int s, i, j;
 177     const uint32_t *sq = ff_square_tab + 256;
 178
 179     s = 0;
 180     for (i = 0; i < w; i++) {
 181         for (j = 0; j < w; j ++) {
 182             s += sq[pix[0]];
 183             pix ++;
 184         }
 185         pix += line_size - w;
 186     }
 187     return s;
 188 }
 189
 190 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 191     switch(type&0xFF){
 192     default:
 193     case FF_CMP_SAD:
 194         return lambda>>FF_LAMBDA_SHIFT;
 195     case FF_CMP_DCT:
 196         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 197     case FF_CMP_W53:
 198         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 199     case FF_CMP_W97:
 200         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 201     case FF_CMP_SATD:
 202     case FF_CMP_DCT264:
 203         return (2*lambda)>>FF_LAMBDA_SHIFT;
 204     case FF_CMP_RD:
 205     case FF_CMP_PSNR:
 206     case FF_CMP_SSE:
 207     case FF_CMP_NSSE:
 208         return lambda2>>FF_LAMBDA_SHIFT;
 209     case FF_CMP_BIT:
 210         return 1;
 211     }
 212 }
 213
 214 //FIXME copy&paste
 215 #define P_LEFT P[1]
 216 #define P_TOP P[2]
 217 #define P_TOPRIGHT P[3]
 218 #define P_MEDIAN P[4]
 219 #define P_MV1 P[9]
 220 #define FLAG_QPEL   1 //must be 1
 221
 222 static int encode_q_branch(SnowContext *s, int level, int x, int y){
 223     uint8_t p_buffer[1024];
 224     uint8_t i_buffer[1024];
 225     uint8_t p_state[sizeof(s->block_state)];
 226     uint8_t i_state[sizeof(s->block_state)];
 227     RangeCoder pc, ic;
 228     uint8_t *pbbak= s->c.bytestream;
 229     uint8_t *pbbak_start= s->c.bytestream_start;
 230     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
 231     const int w= s->b_width  << s->block_max_depth;
 232     const int h= s->b_height << s->block_max_depth;
 233     const int rem_depth= s->block_max_depth - level;
 234     const int index= (x + y*w) << rem_depth;
 235     const int block_w= 1<<(LOG2_MB_SIZE - level);
 236     int trx= (x+1)<<rem_depth;
 237     int try= (y+1)<<rem_depth;
 238     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 239     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 240     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
 241     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
 242     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 243     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 244     int pl = left->color[0];
 245     int pcb= left->color[1];
 246     int pcr= left->color[2];
 247     int pmx, pmy;
 248     int mx=0, my=0;
 249     int l,cr,cb;
 250     const int stride= s->current_picture->linesize[0];
 251     const int uvstride= s->current_picture->linesize[1];
 252     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
 253                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
 254                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
 255     int P[10][2];
 256     int16_t last_mv[3][2];
 257     int qpel= !!(s->avctx->flags & AV_CODEC_FLAG_QPEL); //unused
 258     const int shift= 1+qpel;
 259     MotionEstContext *c= &s->m.me;
 260     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 261     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 262     int my_context= av_log2(2*FFABS(left->my - top->my));
 263     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 264     int ref, best_ref, ref_score, ref_mx, ref_my;
 265
 266     av_assert0(sizeof(s->block_state) >= 256);
 267     if(s->keyframe){
 268         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 269         return 0;
 270     }
 271
 272 //    clip predictors / edge ?
 273
 274     P_LEFT[0]= left->mx;
 275     P_LEFT[1]= left->my;
 276     P_TOP [0]= top->mx;
 277     P_TOP [1]= top->my;
 278     P_TOPRIGHT[0]= tr->mx;
 279     P_TOPRIGHT[1]= tr->my;
 280
 281     last_mv[0][0]= s->block[index].mx;
 282     last_mv[0][1]= s->block[index].my;
 283     last_mv[1][0]= right->mx;
 284     last_mv[1][1]= right->my;
 285     last_mv[2][0]= bottom->mx;
 286     last_mv[2][1]= bottom->my;
 287
 288     s->m.mb_stride=2;
 289     s->m.mb_x=
 290     s->m.mb_y= 0;
 291     c->skip= 0;
 292
 293     av_assert1(c->  stride ==   stride);
 294     av_assert1(c->uvstride == uvstride);
 295
 296     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
 297     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
 298     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
 299     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_DMV;
 300
 301     c->xmin = - x*block_w - 16+3;
 302     c->ymin = - y*block_w - 16+3;
 303     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 304     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 305
 306     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
 307     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
 308     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
 309     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
 310     if(P_TOPRIGHT[0] < (c->xmin * (1<<shift))) P_TOPRIGHT[0]= (c->xmin * (1<<shift));
 311     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
 312     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
 313
 314     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
 315     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 316
 317     if (!y) {
 318         c->pred_x= P_LEFT[0];
 319         c->pred_y= P_LEFT[1];
 320     } else {
 321         c->pred_x = P_MEDIAN[0];
 322         c->pred_y = P_MEDIAN[1];
 323     }
 324
 325     score= INT_MAX;
 326     best_ref= 0;
 327     for(ref=0; ref<s->ref_frames; ref++){
 328         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
 329
 330         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
 331                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
 332
 333         av_assert2(ref_mx >= c->xmin);
 334         av_assert2(ref_mx <= c->xmax);
 335         av_assert2(ref_my >= c->ymin);
 336         av_assert2(ref_my <= c->ymax);
 337
 338         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
 339         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
 340         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
 341         if(s->ref_mvs[ref]){
 342             s->ref_mvs[ref][index][0]= ref_mx;
 343             s->ref_mvs[ref][index][1]= ref_my;
 344             s->ref_scores[ref][index]= ref_score;
 345         }
 346         if(score > ref_score){
 347             score= ref_score;
 348             best_ref= ref;
 349             mx= ref_mx;
 350             my= ref_my;
 351         }
 352     }
 353     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
 354
 355   //  subpel search
 356     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
 357     pc= s->c;
 358     pc.bytestream_start=
 359     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
 360     memcpy(p_state, s->block_state, sizeof(s->block_state));
 361
 362     if(level!=s->block_max_depth)
 363         put_rac(&pc, &p_state[4 + s_context], 1);
 364     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
 365     if(s->ref_frames > 1)
 366         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
 367     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
 368     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
 369     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
 370     p_len= pc.bytestream - pc.bytestream_start;
 371     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
 372
 373     block_s= block_w*block_w;
 374     sum = pix_sum(current_data[0], stride, block_w, block_w);
 375     l= (sum + block_s/2)/block_s;
 376     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
 377
 378     if (s->nb_planes > 2) {
 379         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
 380         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 381         cb= (sum + block_s/2)/block_s;
 382     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
 383         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 384         cr= (sum + block_s/2)/block_s;
 385     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
 386     }else
 387         cb = cr = 0;
 388
 389     ic= s->c;
 390     ic.bytestream_start=
 391     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
 392     memcpy(i_state, s->block_state, sizeof(s->block_state));
 393     if(level!=s->block_max_depth)
 394         put_rac(&ic, &i_state[4 + s_context], 1);
 395     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
 396     put_symbol(&ic, &i_state[32],  l-pl , 1);
 397     if (s->nb_planes > 2) {
 398         put_symbol(&ic, &i_state[64], cb-pcb, 1);
 399         put_symbol(&ic, &i_state[96], cr-pcr, 1);
 400     }
 401     i_len= ic.bytestream - ic.bytestream_start;
 402     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
 403
 404     av_assert1(iscore < 255*255*256 + s->lambda2*10);
 405     av_assert1(iscore >= 0);
 406     av_assert1(l>=0 && l<=255);
 407     av_assert1(pl>=0 && pl<=255);
 408
 409     if(level==0){
 410         int varc= iscore >> 8;
 411         int vard= score >> 8;
 412         if (vard <= 64 || vard < varc)
 413             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
 414         else
 415             c->scene_change_score+= s->m.qscale;
 416     }
 417
 418     if(level!=s->block_max_depth){
 419         put_rac(&s->c, &s->block_state[4 + s_context], 0);
 420         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
 421         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
 422         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
 423         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
 424         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
 425
 426         if(score2 < score && score2 < iscore)
 427             return score2;
 428     }
 429
 430     if(iscore < score){
 431         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 432         memcpy(pbbak, i_buffer, i_len);
 433         s->c= ic;
 434         s->c.bytestream_start= pbbak_start;
 435         s->c.bytestream= pbbak + i_len;
 436         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
 437         memcpy(s->block_state, i_state, sizeof(s->block_state));
 438         return iscore;
 439     }else{
 440         memcpy(pbbak, p_buffer, p_len);
 441         s->c= pc;
 442         s->c.bytestream_start= pbbak_start;
 443         s->c.bytestream= pbbak + p_len;
 444         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
 445         memcpy(s->block_state, p_state, sizeof(s->block_state));
 446         return score;
 447     }
 448 }
 449
 450 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
 451     const int w= s->b_width  << s->block_max_depth;
 452     const int rem_depth= s->block_max_depth - level;
 453     const int index= (x + y*w) << rem_depth;
 454     int trx= (x+1)<<rem_depth;
 455     BlockNode *b= &s->block[index];
 456     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 457     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 458     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 459     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 460     int pl = left->color[0];
 461     int pcb= left->color[1];
 462     int pcr= left->color[2];
 463     int pmx, pmy;
 464     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 465     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
 466     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
 467     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 468
 469     if(s->keyframe){
 470         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 471         return;
 472     }
 473
 474     if(level!=s->block_max_depth){
 475         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
 476             put_rac(&s->c, &s->block_state[4 + s_context], 1);
 477         }else{
 478             put_rac(&s->c, &s->block_state[4 + s_context], 0);
 479             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
 480             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
 481             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
 482             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
 483             return;
 484         }
 485     }
 486     if(b->type & BLOCK_INTRA){
 487         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 488         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
 489         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
 490         if (s->nb_planes > 2) {
 491             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
 492             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
 493         }
 494         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
 495     }else{
 496         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
 497         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
 498         if(s->ref_frames > 1)
 499             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
 500         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
 501         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
 502         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
 503     }
 504 }
 505
 506 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
 507     int i, x2, y2;
 508     Plane *p= &s->plane[plane_index];
 509     const int block_size = MB_SIZE >> s->block_max_depth;
 510     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 511     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 512     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 513     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 514     const int ref_stride= s->current_picture->linesize[plane_index];
 515     uint8_t *src= s-> input_picture->data[plane_index];
 516     IDWTELEM *dst= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
 517     const int b_stride = s->b_width << s->block_max_depth;
 518     const int w= p->width;
 519     const int h= p->height;
 520     int index= mb_x + mb_y*b_stride;
 521     BlockNode *b= &s->block[index];
 522     BlockNode backup= *b;
 523     int ab=0;
 524     int aa=0;
 525
 526     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
 527
 528     b->type|= BLOCK_INTRA;
 529     b->color[plane_index]= 0;
 530     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
 531
 532     for(i=0; i<4; i++){
 533         int mb_x2= mb_x + (i &1) - 1;
 534         int mb_y2= mb_y + (i>>1) - 1;
 535         int x= block_w*mb_x2 + block_w/2;
 536         int y= block_h*mb_y2 + block_h/2;
 537
 538         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
 539                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
 540
 541         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
 542             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
 543                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
 544                 int obmc_v= obmc[index];
 545                 int d;
 546                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
 547                 if(x<0) obmc_v += obmc[index + block_w];
 548                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
 549                 if(x+block_w>w) obmc_v += obmc[index - block_w];
 550                 //FIXME precalculate this or simplify it somehow else
 551
 552                 d = -dst[index] + (1<<(FRAC_BITS-1));
 553                 dst[index] = d;
 554                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
 555                 aa += obmc_v * obmc_v; //FIXME precalculate this
 556             }
 557         }
 558     }
 559     *b= backup;
 560
 561     return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
 562 }
 563
 564 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
 565     const int b_stride = s->b_width << s->block_max_depth;
 566     const int b_height = s->b_height<< s->block_max_depth;
 567     int index= x + y*b_stride;
 568     const BlockNode *b     = &s->block[index];
 569     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 570     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
 571     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
 572     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
 573     int dmx, dmy;
 574 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 575 //  int my_context= av_log2(2*FFABS(left->my - top->my));
 576
 577     if(x<0 || x>=b_stride || y>=b_height)
 578         return 0;
 579 /*
 580 1            0      0
 581 01X          1-2    1
 582 001XX        3-6    2-3
 583 0001XXX      7-14   4-7
 584 00001XXXX   15-30   8-15
 585 */
 586 //FIXME try accurate rate
 587 //FIXME intra and inter predictors if surrounding blocks are not the same type
 588     if(b->type & BLOCK_INTRA){
 589         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
 590                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
 591                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
 592     }else{
 593         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
 594         dmx-= b->mx;
 595         dmy-= b->my;
 596         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
 597                     + av_log2(2*FFABS(dmy))
 598                     + av_log2(2*b->ref));
 599     }
 600 }
 601
 602 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
 603     Plane *p= &s->plane[plane_index];
 604     const int block_size = MB_SIZE >> s->block_max_depth;
 605     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 606     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 607     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 608     const int ref_stride= s->current_picture->linesize[plane_index];
 609     uint8_t *dst= s->current_picture->data[plane_index];
 610     uint8_t *src= s->  input_picture->data[plane_index];
 611     IDWTELEM *pred= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
 612     uint8_t *cur = s->scratchbuf;
 613     uint8_t *tmp = s->emu_edge_buffer;
 614     const int b_stride = s->b_width << s->block_max_depth;
 615     const int b_height = s->b_height<< s->block_max_depth;
 616     const int w= p->width;
 617     const int h= p->height;
 618     int distortion;
 619     int rate= 0;
 620     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 621     int sx= block_w*mb_x - block_w/2;
 622     int sy= block_h*mb_y - block_h/2;
 623     int x0= FFMAX(0,-sx);
 624     int y0= FFMAX(0,-sy);
 625     int x1= FFMIN(block_w*2, w-sx);
 626     int y1= FFMIN(block_h*2, h-sy);
 627     int i,x,y;
 628
 629     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
 630
 631     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);
 632
 633     for(y=y0; y<y1; y++){
 634         const uint8_t *obmc1= obmc_edged[y];
 635         const IDWTELEM *pred1 = pred + y*obmc_stride;
 636         uint8_t *cur1 = cur + y*ref_stride;
 637         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
 638         for(x=x0; x<x1; x++){
 639 #if FRAC_BITS >= LOG2_OBMC_MAX
 640             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
 641 #else
 642             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
 643 #endif
 644             v = (v + pred1[x]) >> FRAC_BITS;
 645             if(v&(~255)) v= ~(v>>31);
 646             dst1[x] = v;
 647         }
 648     }
 649
 650     /* copy the regions where obmc[] = (uint8_t)256 */
 651     if(LOG2_OBMC_MAX == 8
 652         && (mb_x == 0 || mb_x == b_stride-1)
 653         && (mb_y == 0 || mb_y == b_height-1)){
 654         if(mb_x == 0)
 655             x1 = block_w;
 656         else
 657             x0 = block_w;
 658         if(mb_y == 0)
 659             y1 = block_h;
 660         else
 661             y0 = block_h;
 662         for(y=y0; y<y1; y++)
 663             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
 664     }
 665
 666     if(block_w==16){
 667         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
 668         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
 669         /* FIXME cmps overlap but do not cover the wavelet's whole support.
 670          * So improving the score of one block is not strictly guaranteed
 671          * to improve the score of the whole frame, thus iterative motion
 672          * estimation does not always converge. */
 673         if(s->avctx->me_cmp == FF_CMP_W97)
 674             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 675         else if(s->avctx->me_cmp == FF_CMP_W53)
 676             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 677         else{
 678             distortion = 0;
 679             for(i=0; i<4; i++){
 680                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
 681                 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
 682             }
 683         }
 684     }else{
 685         av_assert2(block_w==8);
 686         distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
 687     }
 688
 689     if(plane_index==0){
 690         for(i=0; i<4; i++){
 691 /* ..RRr
 692  * .RXx.
 693  * rxx..
 694  */
 695             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
 696         }
 697         if(mb_x == b_stride-2)
 698             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
 699     }
 700     return distortion + rate*penalty_factor;
 701 }
 702
 703 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
 704     int i, y2;
 705     Plane *p= &s->plane[plane_index];
 706     const int block_size = MB_SIZE >> s->block_max_depth;
 707     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 708     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 709     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 710     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 711     const int ref_stride= s->current_picture->linesize[plane_index];
 712     uint8_t *dst= s->current_picture->data[plane_index];
 713     uint8_t *src= s-> input_picture->data[plane_index];
 714     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
 715     // const has only been removed from zero_dst to suppress a warning
 716     static IDWTELEM zero_dst[4096]; //FIXME
 717     const int b_stride = s->b_width << s->block_max_depth;
 718     const int w= p->width;
 719     const int h= p->height;
 720     int distortion= 0;
 721     int rate= 0;
 722     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 723
 724     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
 725
 726     for(i=0; i<9; i++){
 727         int mb_x2= mb_x + (i%3) - 1;
 728         int mb_y2= mb_y + (i/3) - 1;
 729         int x= block_w*mb_x2 + block_w/2;
 730         int y= block_h*mb_y2 + block_h/2;
 731
 732         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
 733                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
 734
 735         //FIXME find a cleaner/simpler way to skip the outside stuff
 736         for(y2= y; y2<0; y2++)
 737             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 738         for(y2= h; y2<y+block_h; y2++)
 739             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 740         if(x<0){
 741             for(y2= y; y2<y+block_h; y2++)
 742                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
 743         }
 744         if(x+block_w > w){
 745             for(y2= y; y2<y+block_h; y2++)
 746                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
 747         }
 748
 749         av_assert1(block_w== 8 || block_w==16);
 750         distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
 751     }
 752
 753     if(plane_index==0){
 754         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
 755         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
 756
 757 /* ..RRRr
 758  * .RXXx.
 759  * .RXXx.
 760  * rxxx.
 761  */
 762         if(merged)
 763             rate = get_block_bits(s, mb_x, mb_y, 2);
 764         for(i=merged?4:0; i<9; i++){
 765             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
 766             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
 767         }
 768     }
 769     return distortion + rate*penalty_factor;
 770 }
 771
 772 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 773     const int w= b->width;
 774     const int h= b->height;
 775     int x, y;
 776
 777     if(1){
 778         int run=0;
 779         int *runs = s->run_buffer;
 780         int run_index=0;
 781         int max_index;
 782
 783         for(y=0; y<h; y++){
 784             for(x=0; x<w; x++){
 785                 int v, p=0;
 786                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 787                 v= src[x + y*stride];
 788
 789                 if(y){
 790                     t= src[x + (y-1)*stride];
 791                     if(x){
 792                         lt= src[x - 1 + (y-1)*stride];
 793                     }
 794                     if(x + 1 < w){
 795                         rt= src[x + 1 + (y-1)*stride];
 796                     }
 797                 }
 798                 if(x){
 799                     l= src[x - 1 + y*stride];
 800                     /*if(x > 1){
 801                         if(orientation==1) ll= src[y + (x-2)*stride];
 802                         else               ll= src[x - 2 + y*stride];
 803                     }*/
 804                 }
 805                 if(parent){
 806                     int px= x>>1;
 807                     int py= y>>1;
 808                     if(px<b->parent->width && py<b->parent->height)
 809                         p= parent[px + py*2*stride];
 810                 }
 811                 if(!(/*ll|*/l|lt|t|rt|p)){
 812                     if(v){
 813                         runs[run_index++]= run;
 814                         run=0;
 815                     }else{
 816                         run++;
 817                     }
 818                 }
 819             }
 820         }
 821         max_index= run_index;
 822         runs[run_index++]= run;
 823         run_index=0;
 824         run= runs[run_index++];
 825
 826         put_symbol2(&s->c, b->state[30], max_index, 0);
 827         if(run_index <= max_index)
 828             put_symbol2(&s->c, b->state[1], run, 3);
 829
 830         for(y=0; y<h; y++){
 831             if(s->c.bytestream_end - s->c.bytestream < w*40){
 832                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 833                 return AVERROR(ENOMEM);
 834             }
 835             for(x=0; x<w; x++){
 836                 int v, p=0;
 837                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 838                 v= src[x + y*stride];
 839
 840                 if(y){
 841                     t= src[x + (y-1)*stride];
 842                     if(x){
 843                         lt= src[x - 1 + (y-1)*stride];
 844                     }
 845                     if(x + 1 < w){
 846                         rt= src[x + 1 + (y-1)*stride];
 847                     }
 848                 }
 849                 if(x){
 850                     l= src[x - 1 + y*stride];
 851                     /*if(x > 1){
 852                         if(orientation==1) ll= src[y + (x-2)*stride];
 853                         else               ll= src[x - 2 + y*stride];
 854                     }*/
 855                 }
 856                 if(parent){
 857                     int px= x>>1;
 858                     int py= y>>1;
 859                     if(px<b->parent->width && py<b->parent->height)
 860                         p= parent[px + py*2*stride];
 861                 }
 862                 if(/*ll|*/l|lt|t|rt|p){
 863                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 864
 865                     put_rac(&s->c, &b->state[0][context], !!v);
 866                 }else{
 867                     if(!run){
 868                         run= runs[run_index++];
 869
 870                         if(run_index <= max_index)
 871                             put_symbol2(&s->c, b->state[1], run, 3);
 872                         av_assert2(v);
 873                     }else{
 874                         run--;
 875                         av_assert2(!v);
 876                     }
 877                 }
 878                 if(v){
 879                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 880                     int l2= 2*FFABS(l) + (l<0);
 881                     int t2= 2*FFABS(t) + (t<0);
 882
 883                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
 884                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
 885                 }
 886             }
 887         }
 888     }
 889     return 0;
 890 }
 891
 892 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 893 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
 894 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
 895     return encode_subband_c0run(s, b, src, parent, stride, orientation);
 896 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
 897 }
 898
 899 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){
 900     const int b_stride= s->b_width << s->block_max_depth;
 901     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 902     BlockNode backup= *block;
 903     unsigned value;
 904     int rd, index;
 905
 906     av_assert2(mb_x>=0 && mb_y>=0);
 907     av_assert2(mb_x<b_stride);
 908
 909     if(intra){
 910         block->color[0] = p[0];
 911         block->color[1] = p[1];
 912         block->color[2] = p[2];
 913         block->type |= BLOCK_INTRA;
 914     }else{
 915         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
 916         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
 917         if(s->me_cache[index] == value)
 918             return 0;
 919         s->me_cache[index]= value;
 920
 921         block->mx= p[0];
 922         block->my= p[1];
 923         block->type &= ~BLOCK_INTRA;
 924     }
 925
 926     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged) + s->intra_penalty * !!intra;
 927
 928 //FIXME chroma
 929     if(rd < *best_rd){
 930         *best_rd= rd;
 931         return 1;
 932     }else{
 933         *block= backup;
 934         return 0;
 935     }
 936 }
 937
 938 /* special case for int[2] args we discard afterwards,
 939  * fixes compilation problem with gcc 2.95 */
 940 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){
 941     int p[2] = {p0, p1};
 942     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
 943 }
 944
 945 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){
 946     const int b_stride= s->b_width << s->block_max_depth;
 947     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 948     BlockNode backup[4];
 949     unsigned value;
 950     int rd, index;
 951
 952     /* We don't initialize backup[] during variable declaration, because
 953      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
 954      * 'int16_t'". */
 955     backup[0] = block[0];
 956     backup[1] = block[1];
 957     backup[2] = block[b_stride];
 958     backup[3] = block[b_stride + 1];
 959
 960     av_assert2(mb_x>=0 && mb_y>=0);
 961     av_assert2(mb_x<b_stride);
 962     av_assert2(((mb_x|mb_y)&1) == 0);
 963
 964     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
 965     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
 966     if(s->me_cache[index] == value)
 967         return 0;
 968     s->me_cache[index]= value;
 969
 970     block->mx= p0;
 971     block->my= p1;
 972     block->ref= ref;
 973     block->type &= ~BLOCK_INTRA;
 974     block[1]= block[b_stride]= block[b_stride+1]= *block;
 975
 976     rd= get_4block_rd(s, mb_x, mb_y, 0);
 977
 978 //FIXME chroma
 979     if(rd < *best_rd){
 980         *best_rd= rd;
 981         return 1;
 982     }else{
 983         block[0]= backup[0];
 984         block[1]= backup[1];
 985         block[b_stride]= backup[2];
 986         block[b_stride+1]= backup[3];
 987         return 0;
 988     }
 989 }
 990
 991 static void iterative_me(SnowContext *s){
 992     int pass, mb_x, mb_y;
 993     const int b_width = s->b_width  << s->block_max_depth;
 994     const int b_height= s->b_height << s->block_max_depth;
 995     const int b_stride= b_width;
 996     int color[3];
 997
 998     {
 999         RangeCoder r = s->c;
1000         uint8_t state[sizeof(s->block_state)];
1001         memcpy(state, s->block_state, sizeof(s->block_state));
1002         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1003             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1004                 encode_q_branch(s, 0, mb_x, mb_y);
1005         s->c = r;
1006         memcpy(s->block_state, state, sizeof(s->block_state));
1007     }
1008
1009     for(pass=0; pass<25; pass++){
1010         int change= 0;
1011
1012         for(mb_y= 0; mb_y<b_height; mb_y++){
1013             for(mb_x= 0; mb_x<b_width; mb_x++){
1014                 int dia_change, i, j, ref;
1015                 int best_rd= INT_MAX, ref_rd;
1016                 BlockNode backup, ref_b;
1017                 const int index= mb_x + mb_y * b_stride;
1018                 BlockNode *block= &s->block[index];
1019                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1020                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1021                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1022                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1023                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1024                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1025                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1026                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1027                 const int b_w= (MB_SIZE >> s->block_max_depth);
1028                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1029
1030                 if(pass && (block->type & BLOCK_OPT))
1031                     continue;
1032                 block->type |= BLOCK_OPT;
1033
1034                 backup= *block;
1035
1036                 if(!s->me_cache_generation)
1037                     memset(s->me_cache, 0, sizeof(s->me_cache));
1038                 s->me_cache_generation += 1<<22;
1039
1040                 //FIXME precalculate
1041                 {
1042                     int x, y;
1043                     for (y = 0; y < b_w * 2; y++)
1044                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1045                     if(mb_x==0)
1046                         for(y=0; y<b_w*2; y++)
1047                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1048                     if(mb_x==b_stride-1)
1049                         for(y=0; y<b_w*2; y++)
1050                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1051                     if(mb_y==0){
1052                         for(x=0; x<b_w*2; x++)
1053                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1054                         for(y=1; y<b_w; y++)
1055                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1056                     }
1057                     if(mb_y==b_height-1){
1058                         for(x=0; x<b_w*2; x++)
1059                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1060                         for(y=b_w; y<b_w*2-1; y++)
1061                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1062                     }
1063                 }
1064
1065                 //skip stuff outside the picture
1066                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1067                     uint8_t *src= s->  input_picture->data[0];
1068                     uint8_t *dst= s->current_picture->data[0];
1069                     const int stride= s->current_picture->linesize[0];
1070                     const int block_w= MB_SIZE >> s->block_max_depth;
1071                     const int block_h= MB_SIZE >> s->block_max_depth;
1072                     const int sx= block_w*mb_x - block_w/2;
1073                     const int sy= block_h*mb_y - block_h/2;
1074                     const int w= s->plane[0].width;
1075                     const int h= s->plane[0].height;
1076                     int y;
1077
1078                     for(y=sy; y<0; y++)
1079                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1080                     for(y=h; y<sy+block_h*2; y++)
1081                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1082                     if(sx<0){
1083                         for(y=sy; y<sy+block_h*2; y++)
1084                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1085                     }
1086                     if(sx+block_w*2 > w){
1087                         for(y=sy; y<sy+block_h*2; y++)
1088                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1089                     }
1090                 }
1091
1092                 // intra(black) = neighbors' contribution to the current block
1093                 for(i=0; i < s->nb_planes; i++)
1094                     color[i]= get_dc(s, mb_x, mb_y, i);
1095
1096                 // get previous score (cannot be cached due to OBMC)
1097                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1098                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1099                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1100                 }else
1101                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1102
1103                 ref_b= *block;
1104                 ref_rd= best_rd;
1105                 for(ref=0; ref < s->ref_frames; ref++){
1106                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1107                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1108                         continue;
1109                     block->ref= ref;
1110                     best_rd= INT_MAX;
1111
1112                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1113                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1114                     if(tb)
1115                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1116                     if(lb)
1117                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1118                     if(rb)
1119                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1120                     if(bb)
1121                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1122
1123                     /* fullpel ME */
1124                     //FIXME avoid subpel interpolation / round to nearest integer
1125                     do{
1126                         int newx = block->mx;
1127                         int newy = block->my;
1128                         int dia_size = s->iterative_dia_size ? s->iterative_dia_size : FFMAX(s->avctx->dia_size, 1);
1129                         dia_change=0;
1130                         for(i=0; i < dia_size; i++){
1131                             for(j=0; j<i; j++){
1132                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
1133                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-4*(i-j), newy-(4*j), obmc_edged, &best_rd);
1134                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-(4*j), newy+4*(i-j), obmc_edged, &best_rd);
1135                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+(4*j), newy-4*(i-j), obmc_edged, &best_rd);
1136                             }
1137                         }
1138                     }while(dia_change);
1139                     /* subpel ME */
1140                     do{
1141                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1142                         dia_change=0;
1143                         for(i=0; i<8; i++)
1144                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1145                     }while(dia_change);
1146                     //FIXME or try the standard 2 pass qpel or similar
1147
1148                     mvr[0][0]= block->mx;
1149                     mvr[0][1]= block->my;
1150                     if(ref_rd > best_rd){
1151                         ref_rd= best_rd;
1152                         ref_b= *block;
1153                     }
1154                 }
1155                 best_rd= ref_rd;
1156                 *block= ref_b;
1157                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1158                 //FIXME RD style color selection
1159                 if(!same_block(block, &backup)){
1160                     if(tb ) tb ->type &= ~BLOCK_OPT;
1161                     if(lb ) lb ->type &= ~BLOCK_OPT;
1162                     if(rb ) rb ->type &= ~BLOCK_OPT;
1163                     if(bb ) bb ->type &= ~BLOCK_OPT;
1164                     if(tlb) tlb->type &= ~BLOCK_OPT;
1165                     if(trb) trb->type &= ~BLOCK_OPT;
1166                     if(blb) blb->type &= ~BLOCK_OPT;
1167                     if(brb) brb->type &= ~BLOCK_OPT;
1168                     change ++;
1169                 }
1170             }
1171         }
1172         av_log(s->avctx, AV_LOG_DEBUG, "pass:%d changed:%d\n", pass, change);
1173         if(!change)
1174             break;
1175     }
1176
1177     if(s->block_max_depth == 1){
1178         int change= 0;
1179         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1180             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1181                 int i;
1182                 int best_rd, init_rd;
1183                 const int index= mb_x + mb_y * b_stride;
1184                 BlockNode *b[4];
1185
1186                 b[0]= &s->block[index];
1187                 b[1]= b[0]+1;
1188                 b[2]= b[0]+b_stride;
1189                 b[3]= b[2]+1;
1190                 if(same_block(b[0], b[1]) &&
1191                    same_block(b[0], b[2]) &&
1192                    same_block(b[0], b[3]))
1193                     continue;
1194
1195                 if(!s->me_cache_generation)
1196                     memset(s->me_cache, 0, sizeof(s->me_cache));
1197                 s->me_cache_generation += 1<<22;
1198
1199                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1200
1201                 //FIXME more multiref search?
1202                 check_4block_inter(s, mb_x, mb_y,
1203                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1204                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1205
1206                 for(i=0; i<4; i++)
1207                     if(!(b[i]->type&BLOCK_INTRA))
1208                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1209
1210                 if(init_rd != best_rd)
1211                     change++;
1212             }
1213         }
1214         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1215     }
1216 }
1217
1218 static void encode_blocks(SnowContext *s, int search){
1219     int x, y;
1220     int w= s->b_width;
1221     int h= s->b_height;
1222
1223     if(s->motion_est == FF_ME_ITER && !s->keyframe && search)
1224         iterative_me(s);
1225
1226     for(y=0; y<h; y++){
1227         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1228             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1229             return;
1230         }
1231         for(x=0; x<w; x++){
1232             if(s->motion_est == FF_ME_ITER || !search)
1233                 encode_q_branch2(s, 0, x, y);
1234             else
1235                 encode_q_branch (s, 0, x, y);
1236         }
1237     }
1238 }
1239
1240 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1241     const int w= b->width;
1242     const int h= b->height;
1243     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1244     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1245     int x,y, thres1, thres2;
1246
1247     if(s->qlog == LOSSLESS_QLOG){
1248         for(y=0; y<h; y++)
1249             for(x=0; x<w; x++)
1250                 dst[x + y*stride]= src[x + y*stride];
1251         return;
1252     }
1253
1254     bias= bias ? 0 : (3*qmul)>>3;
1255     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1256     thres2= 2*thres1;
1257
1258     if(!bias){
1259         for(y=0; y<h; y++){
1260             for(x=0; x<w; x++){
1261                 int i= src[x + y*stride];
1262
1263                 if((unsigned)(i+thres1) > thres2){
1264                     if(i>=0){
1265                         i<<= QEXPSHIFT;
1266                         i/= qmul; //FIXME optimize
1267                         dst[x + y*stride]=  i;
1268                     }else{
1269                         i= -i;
1270                         i<<= QEXPSHIFT;
1271                         i/= qmul; //FIXME optimize
1272                         dst[x + y*stride]= -i;
1273                     }
1274                 }else
1275                     dst[x + y*stride]= 0;
1276             }
1277         }
1278     }else{
1279         for(y=0; y<h; y++){
1280             for(x=0; x<w; x++){
1281                 int i= src[x + y*stride];
1282
1283                 if((unsigned)(i+thres1) > thres2){
1284                     if(i>=0){
1285                         i<<= QEXPSHIFT;
1286                         i= (i + bias) / qmul; //FIXME optimize
1287                         dst[x + y*stride]=  i;
1288                     }else{
1289                         i= -i;
1290                         i<<= QEXPSHIFT;
1291                         i= (i + bias) / qmul; //FIXME optimize
1292                         dst[x + y*stride]= -i;
1293                     }
1294                 }else
1295                     dst[x + y*stride]= 0;
1296             }
1297         }
1298     }
1299 }
1300
1301 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1302     const int w= b->width;
1303     const int h= b->height;
1304     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1305     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1306     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1307     int x,y;
1308
1309     if(s->qlog == LOSSLESS_QLOG) return;
1310
1311     for(y=0; y<h; y++){
1312         for(x=0; x<w; x++){
1313             int i= src[x + y*stride];
1314             if(i<0){
1315                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1316             }else if(i>0){
1317                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1318             }
1319         }
1320     }
1321 }
1322
1323 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1324     const int w= b->width;
1325     const int h= b->height;
1326     int x,y;
1327
1328     for(y=h-1; y>=0; y--){
1329         for(x=w-1; x>=0; x--){
1330             int i= x + y*stride;
1331
1332             if(x){
1333                 if(use_median){
1334                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1335                     else  src[i] -= src[i - 1];
1336                 }else{
1337                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1338                     else  src[i] -= src[i - 1];
1339                 }
1340             }else{
1341                 if(y) src[i] -= src[i - stride];
1342             }
1343         }
1344     }
1345 }
1346
1347 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1348     const int w= b->width;
1349     const int h= b->height;
1350     int x,y;
1351
1352     for(y=0; y<h; y++){
1353         for(x=0; x<w; x++){
1354             int i= x + y*stride;
1355
1356             if(x){
1357                 if(use_median){
1358                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1359                     else  src[i] += src[i - 1];
1360                 }else{
1361                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1362                     else  src[i] += src[i - 1];
1363                 }
1364             }else{
1365                 if(y) src[i] += src[i - stride];
1366             }
1367         }
1368     }
1369 }
1370
1371 static void encode_qlogs(SnowContext *s){
1372     int plane_index, level, orientation;
1373
1374     for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1375         for(level=0; level<s->spatial_decomposition_count; level++){
1376             for(orientation=level ? 1:0; orientation<4; orientation++){
1377                 if(orientation==2) continue;
1378                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1379             }
1380         }
1381     }
1382 }
1383
1384 static void encode_header(SnowContext *s){
1385     int plane_index, i;
1386     uint8_t kstate[32];
1387
1388     memset(kstate, MID_STATE, sizeof(kstate));
1389
1390     put_rac(&s->c, kstate, s->keyframe);
1391     if(s->keyframe || s->always_reset){
1392         ff_snow_reset_contexts(s);
1393         s->last_spatial_decomposition_type=
1394         s->last_qlog=
1395         s->last_qbias=
1396         s->last_mv_scale=
1397         s->last_block_max_depth= 0;
1398         for(plane_index=0; plane_index<2; plane_index++){
1399             Plane *p= &s->plane[plane_index];
1400             p->last_htaps=0;
1401             p->last_diag_mc=0;
1402             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1403         }
1404     }
1405     if(s->keyframe){
1406         put_symbol(&s->c, s->header_state, s->version, 0);
1407         put_rac(&s->c, s->header_state, s->always_reset);
1408         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1409         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1410         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1411         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1412         if (s->nb_planes > 2) {
1413             put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1414             put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1415         }
1416         put_rac(&s->c, s->header_state, s->spatial_scalability);
1417 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1418         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1419
1420         encode_qlogs(s);
1421     }
1422
1423     if(!s->keyframe){
1424         int update_mc=0;
1425         for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1426             Plane *p= &s->plane[plane_index];
1427             update_mc |= p->last_htaps   != p->htaps;
1428             update_mc |= p->last_diag_mc != p->diag_mc;
1429             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1430         }
1431         put_rac(&s->c, s->header_state, update_mc);
1432         if(update_mc){
1433             for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1434                 Plane *p= &s->plane[plane_index];
1435                 put_rac(&s->c, s->header_state, p->diag_mc);
1436                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1437                 for(i= p->htaps/2; i; i--)
1438                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1439             }
1440         }
1441         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1442             put_rac(&s->c, s->header_state, 1);
1443             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1444             encode_qlogs(s);
1445         }else
1446             put_rac(&s->c, s->header_state, 0);
1447     }
1448
1449     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1450     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1451     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1452     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1453     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1454
1455 }
1456
1457 static void update_last_header_values(SnowContext *s){
1458     int plane_index;
1459
1460     if(!s->keyframe){
1461         for(plane_index=0; plane_index<2; plane_index++){
1462             Plane *p= &s->plane[plane_index];
1463             p->last_diag_mc= p->diag_mc;
1464             p->last_htaps  = p->htaps;
1465             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1466         }
1467     }
1468
1469     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1470     s->last_qlog                        = s->qlog;
1471     s->last_qbias                       = s->qbias;
1472     s->last_mv_scale                    = s->mv_scale;
1473     s->last_block_max_depth             = s->block_max_depth;
1474     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1475 }
1476
1477 static int qscale2qlog(int qscale){
1478     return lrint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1479            + 61*QROOT/8; ///< 64 > 60
1480 }
1481
1482 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1483 {
1484     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1485      * FIXME we know exact mv bits at this point,
1486      * but ratecontrol isn't set up to include them. */
1487     uint32_t coef_sum= 0;
1488     int level, orientation, delta_qlog;
1489
1490     for(level=0; level<s->spatial_decomposition_count; level++){
1491         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1492             SubBand *b= &s->plane[0].band[level][orientation];
1493             IDWTELEM *buf= b->ibuf;
1494             const int w= b->width;
1495             const int h= b->height;
1496             const int stride= b->stride;
1497             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1498             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1499             const int qdiv= (1<<16)/qmul;
1500             int x, y;
1501             //FIXME this is ugly
1502             for(y=0; y<h; y++)
1503                 for(x=0; x<w; x++)
1504                     buf[x+y*stride]= b->buf[x+y*stride];
1505             if(orientation==0)
1506                 decorrelate(s, b, buf, stride, 1, 0);
1507             for(y=0; y<h; y++)
1508                 for(x=0; x<w; x++)
1509                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1510         }
1511     }
1512
1513     /* ugly, ratecontrol just takes a sqrt again */
1514     av_assert0(coef_sum < INT_MAX);
1515     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1516
1517     if(pict->pict_type == AV_PICTURE_TYPE_I){
1518         s->m.current_picture.mb_var_sum= coef_sum;
1519         s->m.current_picture.mc_mb_var_sum= 0;
1520     }else{
1521         s->m.current_picture.mc_mb_var_sum= coef_sum;
1522         s->m.current_picture.mb_var_sum= 0;
1523     }
1524
1525     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1526     if (pict->quality < 0)
1527         return INT_MIN;
1528     s->lambda= pict->quality * 3/2;
1529     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1530     s->qlog+= delta_qlog;
1531     return delta_qlog;
1532 }
1533
1534 static void calculate_visual_weight(SnowContext *s, Plane *p){
1535     int width = p->width;
1536     int height= p->height;
1537     int level, orientation, x, y;
1538
1539     for(level=0; level<s->spatial_decomposition_count; level++){
1540         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1541             SubBand *b= &p->band[level][orientation];
1542             IDWTELEM *ibuf= b->ibuf;
1543             int64_t error=0;
1544
1545             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1546             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1547             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1548             for(y=0; y<height; y++){
1549                 for(x=0; x<width; x++){
1550                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1551                     error += d*d;
1552                 }
1553             }
1554
1555             b->qlog= (int)(QROOT * log2(352256.0/sqrt(error)) + 0.5);
1556         }
1557     }
1558 }
1559
1560 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1561                         const AVFrame *pict, int *got_packet)
1562 {
1563     SnowContext *s = avctx->priv_data;
1564     RangeCoder * const c= &s->c;
1565     AVFrame *pic;
1566     const int width= s->avctx->width;
1567     const int height= s->avctx->height;
1568     int level, orientation, plane_index, i, y, ret;
1569     uint8_t rc_header_bak[sizeof(s->header_state)];
1570     uint8_t rc_block_bak[sizeof(s->block_state)];
1571
1572     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)
1573         return ret;
1574
1575     ff_init_range_encoder(c, pkt->data, pkt->size);
1576     ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1577
1578     for(i=0; i < s->nb_planes; i++){
1579         int hshift= i ? s->chroma_h_shift : 0;
1580         int vshift= i ? s->chroma_v_shift : 0;
1581         for(y=0; y<AV_CEIL_RSHIFT(height, vshift); y++)
1582             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1583                    &pict->data[i][y * pict->linesize[i]],
1584                    AV_CEIL_RSHIFT(width, hshift));
1585         s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1586                                 AV_CEIL_RSHIFT(width, hshift), AV_CEIL_RSHIFT(height, vshift),
1587                                 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1588                                 EDGE_TOP | EDGE_BOTTOM);
1589
1590     }
1591     emms_c();
1592     pic = s->input_picture;
1593     pic->pict_type = pict->pict_type;
1594     pic->quality = pict->quality;
1595
1596     s->m.picture_number= avctx->frame_number;
1597     if(avctx->flags&AV_CODEC_FLAG_PASS2){
1598         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1599         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1600         if(!(avctx->flags&AV_CODEC_FLAG_QSCALE)) {
1601             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1602             if (pic->quality < 0)
1603                 return -1;
1604         }
1605     }else{
1606         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1607         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1608     }
1609
1610     if(s->pass1_rc && avctx->frame_number == 0)
1611         pic->quality = 2*FF_QP2LAMBDA;
1612     if (pic->quality) {
1613         s->qlog   = qscale2qlog(pic->quality);
1614         s->lambda = pic->quality * 3/2;
1615     }
1616     if (s->qlog < 0 || (!pic->quality && (avctx->flags & AV_CODEC_FLAG_QSCALE))) {
1617         s->qlog= LOSSLESS_QLOG;
1618         s->lambda = 0;
1619     }//else keep previous frame's qlog until after motion estimation
1620
1621     if (s->current_picture->data[0]) {
1622         int w = s->avctx->width;
1623         int h = s->avctx->height;
1624
1625         s->mpvencdsp.draw_edges(s->current_picture->data[0],
1626                                 s->current_picture->linesize[0], w   , h   ,
1627                                 EDGE_WIDTH  , EDGE_WIDTH  , EDGE_TOP | EDGE_BOTTOM);
1628         if (s->current_picture->data[2]) {
1629             s->mpvencdsp.draw_edges(s->current_picture->data[1],
1630                                     s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1631                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1632             s->mpvencdsp.draw_edges(s->current_picture->data[2],
1633                                     s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1634                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1635         }
1636         emms_c();
1637     }
1638
1639     ff_snow_frame_start(s);
1640
1641     s->m.current_picture_ptr= &s->m.current_picture;
1642     s->m.current_picture.f = s->current_picture;
1643     s->m.current_picture.f->pts = pict->pts;
1644     if(pic->pict_type == AV_PICTURE_TYPE_P){
1645         int block_width = (width +15)>>4;
1646         int block_height= (height+15)>>4;
1647         int stride= s->current_picture->linesize[0];
1648
1649         av_assert0(s->current_picture->data[0]);
1650         av_assert0(s->last_picture[0]->data[0]);
1651
1652         s->m.avctx= s->avctx;
1653         s->m.   last_picture.f = s->last_picture[0];
1654         s->m.    new_picture.f = s->input_picture;
1655         s->m.   last_picture_ptr= &s->m.   last_picture;
1656         s->m.linesize = stride;
1657         s->m.uvlinesize= s->current_picture->linesize[1];
1658         s->m.width = width;
1659         s->m.height= height;
1660         s->m.mb_width = block_width;
1661         s->m.mb_height= block_height;
1662         s->m.mb_stride=   s->m.mb_width+1;
1663         s->m.b8_stride= 2*s->m.mb_width+1;
1664         s->m.f_code=1;
1665         s->m.pict_type = pic->pict_type;
1666         s->m.motion_est= s->motion_est;
1667         s->m.me.scene_change_score=0;
1668         s->m.me.dia_size = avctx->dia_size;
1669         s->m.quarter_sample= (s->avctx->flags & AV_CODEC_FLAG_QPEL)!=0;
1670         s->m.out_format= FMT_H263;
1671         s->m.unrestricted_mv= 1;
1672
1673         s->m.lambda = s->lambda;
1674         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1675         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1676
1677         s->m.mecc= s->mecc; //move
1678         s->m.qdsp= s->qdsp; //move
1679         s->m.hdsp = s->hdsp;
1680         ff_init_me(&s->m);
1681         s->hdsp = s->m.hdsp;
1682         s->mecc= s->m.mecc;
1683     }
1684
1685     if(s->pass1_rc){
1686         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1687         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1688     }
1689
1690 redo_frame:
1691
1692     s->spatial_decomposition_count= 5;
1693
1694     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1695           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1696         s->spatial_decomposition_count--;
1697
1698     if (s->spatial_decomposition_count <= 0) {
1699         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1700         return AVERROR(EINVAL);
1701     }
1702
1703     s->m.pict_type = pic->pict_type;
1704     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1705
1706     ff_snow_common_init_after_header(avctx);
1707
1708     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1709         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1710             calculate_visual_weight(s, &s->plane[plane_index]);
1711         }
1712     }
1713
1714     encode_header(s);
1715     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1716     encode_blocks(s, 1);
1717     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1718
1719     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1720         Plane *p= &s->plane[plane_index];
1721         int w= p->width;
1722         int h= p->height;
1723         int x, y;
1724 //        int bits= put_bits_count(&s->c.pb);
1725
1726         if (!s->memc_only) {
1727             //FIXME optimize
1728             if(pict->data[plane_index]) //FIXME gray hack
1729                 for(y=0; y<h; y++){
1730                     for(x=0; x<w; x++){
1731                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1732                     }
1733                 }
1734             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1735
1736             if(   plane_index==0
1737                && pic->pict_type == AV_PICTURE_TYPE_P
1738                && !(avctx->flags&AV_CODEC_FLAG_PASS2)
1739                && s->m.me.scene_change_score > s->scenechange_threshold){
1740                 ff_init_range_encoder(c, pkt->data, pkt->size);
1741                 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1742                 pic->pict_type= AV_PICTURE_TYPE_I;
1743                 s->keyframe=1;
1744                 s->current_picture->key_frame=1;
1745                 goto redo_frame;
1746             }
1747
1748             if(s->qlog == LOSSLESS_QLOG){
1749                 for(y=0; y<h; y++){
1750                     for(x=0; x<w; x++){
1751                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1752                     }
1753                 }
1754             }else{
1755                 for(y=0; y<h; y++){
1756                     for(x=0; x<w; x++){
1757                         s->spatial_dwt_buffer[y*w + x]= s->spatial_idwt_buffer[y*w + x] * (1 << ENCODER_EXTRA_BITS);
1758                     }
1759                 }
1760             }
1761
1762             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1763
1764             if(s->pass1_rc && plane_index==0){
1765                 int delta_qlog = ratecontrol_1pass(s, pic);
1766                 if (delta_qlog <= INT_MIN)
1767                     return -1;
1768                 if(delta_qlog){
1769                     //reordering qlog in the bitstream would eliminate this reset
1770                     ff_init_range_encoder(c, pkt->data, pkt->size);
1771                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1772                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1773                     encode_header(s);
1774                     encode_blocks(s, 0);
1775                 }
1776             }
1777
1778             for(level=0; level<s->spatial_decomposition_count; level++){
1779                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1780                     SubBand *b= &p->band[level][orientation];
1781
1782                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1783                     if(orientation==0)
1784                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1785                     if (!s->no_bitstream)
1786                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1787                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1788                     if(orientation==0)
1789                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1790                 }
1791             }
1792
1793             for(level=0; level<s->spatial_decomposition_count; level++){
1794                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1795                     SubBand *b= &p->band[level][orientation];
1796
1797                     dequantize(s, b, b->ibuf, b->stride);
1798                 }
1799             }
1800
1801             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1802             if(s->qlog == LOSSLESS_QLOG){
1803                 for(y=0; y<h; y++){
1804                     for(x=0; x<w; x++){
1805                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1806                     }
1807                 }
1808             }
1809             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1810         }else{
1811             //ME/MC only
1812             if(pic->pict_type == AV_PICTURE_TYPE_I){
1813                 for(y=0; y<h; y++){
1814                     for(x=0; x<w; x++){
1815                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1816                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1817                     }
1818                 }
1819             }else{
1820                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1821                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1822             }
1823         }
1824         if(s->avctx->flags&AV_CODEC_FLAG_PSNR){
1825             int64_t error= 0;
1826
1827             if(pict->data[plane_index]) //FIXME gray hack
1828                 for(y=0; y<h; y++){
1829                     for(x=0; x<w; x++){
1830                         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];
1831                         error += d*d;
1832                     }
1833                 }
1834             s->avctx->error[plane_index] += error;
1835             s->encoding_error[plane_index] = error;
1836         }
1837
1838     }
1839     emms_c();
1840
1841     update_last_header_values(s);
1842
1843     ff_snow_release_buffer(avctx);
1844
1845     s->current_picture->coded_picture_number = avctx->frame_number;
1846     s->current_picture->pict_type = pic->pict_type;
1847     s->current_picture->quality = pic->quality;
1848     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1849     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1850     s->m.current_picture.f->display_picture_number =
1851     s->m.current_picture.f->coded_picture_number   = avctx->frame_number;
1852     s->m.current_picture.f->quality                = pic->quality;
1853     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1854     if(s->pass1_rc)
1855         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1856             return -1;
1857     if(avctx->flags&AV_CODEC_FLAG_PASS1)
1858         ff_write_pass1_stats(&s->m);
1859     s->m.last_pict_type = s->m.pict_type;
1860
1861     emms_c();
1862
1863     ff_side_data_set_encoder_stats(pkt, s->current_picture->quality,
1864                                    s->encoding_error,
1865                                    (s->avctx->flags&AV_CODEC_FLAG_PSNR) ? 4 : 0,
1866                                    s->current_picture->pict_type);
1867
1868     pkt->size = ff_rac_terminate(c, 0);
1869     if (s->current_picture->key_frame)
1870         pkt->flags |= AV_PKT_FLAG_KEY;
1871     *got_packet = 1;
1872
1873     return 0;
1874 }
1875
1876 static av_cold int encode_end(AVCodecContext *avctx)
1877 {
1878     SnowContext *s = avctx->priv_data;
1879
1880     ff_snow_common_end(s);
1881     ff_rate_control_uninit(&s->m);
1882     av_frame_free(&s->input_picture);
1883     av_freep(&avctx->stats_out);
1884
1885     return 0;
1886 }
1887
1888 #define OFFSET(x) offsetof(SnowContext, x)
1889 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1890 static const AVOption options[] = {
1891     {"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" },
1892     { "zero", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ZERO }, 0, 0, VE, "motion_est" },
1893     { "epzs", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_EPZS }, 0, 0, VE, "motion_est" },
1894     { "xone", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_XONE }, 0, 0, VE, "motion_est" },
1895     { "iter", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ITER }, 0, 0, VE, "motion_est" },
1896     { "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 },
1897     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1898     { "intra_penalty",  "Penalty for intra blocks in block decission",      OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1899     { "iterative_dia_size",  "Dia size for the iterative ME",          OFFSET(iterative_dia_size), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1900     { "sc_threshold",   "Scene change threshold",                   OFFSET(scenechange_threshold), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, VE },
1901     { "pred",           "Spatial decomposition type",                                OFFSET(pred), AV_OPT_TYPE_INT, { .i64 = 0 }, DWT_97, DWT_53, VE, "pred" },
1902         { "dwt97", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, VE, "pred" },
1903         { "dwt53", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, VE, "pred" },
1904     { "rc_eq", "Set rate control equation. When computing the expression, besides the standard functions "
1905      "defined in the section 'Expression Evaluation', the following functions are available: "
1906      "bits2qp(bits), qp2bits(qp). Also the following constants are available: iTex pTex tex mv "
1907      "fCode iCount mcVar var isI isP isB avgQP qComp avgIITex avgPITex avgPPTex avgBPTex avgTex.",
1908                                                                                   OFFSET(m.rc_eq), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VE },
1909     { NULL },
1910 };
1911
1912 static const AVClass snowenc_class = {
1913     .class_name = "snow encoder",
1914     .item_name  = av_default_item_name,
1915     .option     = options,
1916     .version    = LIBAVUTIL_VERSION_INT,
1917 };
1918
1919 AVCodec ff_snow_encoder = {
1920     .name           = "snow",
1921     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1922     .type           = AVMEDIA_TYPE_VIDEO,
1923     .id             = AV_CODEC_ID_SNOW,
1924     .priv_data_size = sizeof(SnowContext),
1925     .init           = encode_init,
1926     .encode2        = encode_frame,
1927     .close          = encode_end,
1928     .pix_fmts       = (const enum AVPixelFormat[]){
1929         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1930         AV_PIX_FMT_GRAY8,
1931         AV_PIX_FMT_NONE
1932     },
1933     .priv_class     = &snowenc_class,
1934     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
1935                       FF_CODEC_CAP_INIT_CLEANUP,
1936 };