git.sesse.net Git - ffmpeg/blob - libavcodec/snowenc.c

   1 /*
   2  * Copyright (C) 2004 Michael Niedermayer <michaelni@gmx.at>
   3  *
   4  * This file is part of FFmpeg.
   5  *
   6  * FFmpeg is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * FFmpeg is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with FFmpeg; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 #include "libavutil/intmath.h"
  22 #include "libavutil/log.h"
  23 #include "libavutil/opt.h"
  24 #include "avcodec.h"
  25 #include "dsputil.h"
  26 #include "internal.h"
  27 #include "snow_dwt.h"
  28 #include "snow.h"
  29
  30 #include "rangecoder.h"
  31 #include "mathops.h"
  32
  33 #include "mpegvideo.h"
  34 #include "h263.h"
  35
  36 static av_cold int encode_init(AVCodecContext *avctx)
  37 {
  38     SnowContext *s = avctx->priv_data;
  39     int plane_index, ret;
  40     int i;
  41
  42     if(avctx->prediction_method == DWT_97
  43        && (avctx->flags & CODEC_FLAG_QSCALE)
  44        && avctx->global_quality == 0){
  45         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
  46         return -1;
  47     }
  48
  49     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
  50
  51     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
  52     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
  53
  54     for(plane_index=0; plane_index<3; plane_index++){
  55         s->plane[plane_index].diag_mc= 1;
  56         s->plane[plane_index].htaps= 6;
  57         s->plane[plane_index].hcoeff[0]=  40;
  58         s->plane[plane_index].hcoeff[1]= -10;
  59         s->plane[plane_index].hcoeff[2]=   2;
  60         s->plane[plane_index].fast_mc= 1;
  61     }
  62
  63     if ((ret = ff_snow_common_init(avctx)) < 0) {
  64         ff_snow_common_end(avctx->priv_data);
  65         return ret;
  66     }
  67     ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
  68
  69     ff_snow_alloc_blocks(s);
  70
  71     s->version=0;
  72
  73     s->m.avctx   = avctx;
  74     s->m.flags   = avctx->flags;
  75     s->m.bit_rate= avctx->bit_rate;
  76
  77     s->m.me.temp      =
  78     s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
  79     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  80     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
  81     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
  82     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.obmc_scratchpad)
  83         return AVERROR(ENOMEM);
  84
  85     ff_h263_encode_init(&s->m); //mv_penalty
  86
  87     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
  88
  89     if(avctx->flags&CODEC_FLAG_PASS1){
  90         if(!avctx->stats_out)
  91             avctx->stats_out = av_mallocz(256);
  92
  93         if (!avctx->stats_out)
  94             return AVERROR(ENOMEM);
  95     }
  96     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
  97         if(ff_rate_control_init(&s->m) < 0)
  98             return -1;
  99     }
 100     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
 101
 102     switch(avctx->pix_fmt){
 103     case AV_PIX_FMT_YUV444P:
 104 //    case AV_PIX_FMT_YUV422P:
 105     case AV_PIX_FMT_YUV420P:
 106 //    case AV_PIX_FMT_YUV411P:
 107     case AV_PIX_FMT_YUV410P:
 108         s->nb_planes = 3;
 109         s->colorspace_type= 0;
 110         break;
 111     case AV_PIX_FMT_GRAY8:
 112         s->nb_planes = 1;
 113         s->colorspace_type = 1;
 114         break;
 115 /*    case AV_PIX_FMT_RGB32:
 116         s->colorspace= 1;
 117         break;*/
 118     default:
 119         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
 120         return -1;
 121     }
 122     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
 123
 124     ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
 125     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
 126
 127     s->input_picture = av_frame_alloc();
 128     if (!s->input_picture)
 129         return AVERROR(ENOMEM);
 130
 131     if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
 132         return ret;
 133
 134     if(s->avctx->me_method == ME_ITER){
 135         int size= s->b_width * s->b_height << 2*s->block_max_depth;
 136         for(i=0; i<s->max_ref_frames; i++){
 137             s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
 138             s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
 139             if (!s->ref_mvs[i] || !s->ref_scores[i])
 140                 return AVERROR(ENOMEM);
 141         }
 142     }
 143
 144     return 0;
 145 }
 146
 147 //near copy & paste from dsputil, FIXME
 148 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
 149 {
 150     int s, i, j;
 151
 152     s = 0;
 153     for (i = 0; i < h; i++) {
 154         for (j = 0; j < w; j++) {
 155             s += pix[0];
 156             pix ++;
 157         }
 158         pix += line_size - w;
 159     }
 160     return s;
 161 }
 162
 163 //near copy & paste from dsputil, FIXME
 164 static int pix_norm1(uint8_t * pix, int line_size, int w)
 165 {
 166     int s, i, j;
 167     uint32_t *sq = ff_square_tab + 256;
 168
 169     s = 0;
 170     for (i = 0; i < w; i++) {
 171         for (j = 0; j < w; j ++) {
 172             s += sq[pix[0]];
 173             pix ++;
 174         }
 175         pix += line_size - w;
 176     }
 177     return s;
 178 }
 179
 180 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 181     switch(type&0xFF){
 182     default:
 183     case FF_CMP_SAD:
 184         return lambda>>FF_LAMBDA_SHIFT;
 185     case FF_CMP_DCT:
 186         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 187     case FF_CMP_W53:
 188         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 189     case FF_CMP_W97:
 190         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 191     case FF_CMP_SATD:
 192     case FF_CMP_DCT264:
 193         return (2*lambda)>>FF_LAMBDA_SHIFT;
 194     case FF_CMP_RD:
 195     case FF_CMP_PSNR:
 196     case FF_CMP_SSE:
 197     case FF_CMP_NSSE:
 198         return lambda2>>FF_LAMBDA_SHIFT;
 199     case FF_CMP_BIT:
 200         return 1;
 201     }
 202 }
 203
 204 //FIXME copy&paste
 205 #define P_LEFT P[1]
 206 #define P_TOP P[2]
 207 #define P_TOPRIGHT P[3]
 208 #define P_MEDIAN P[4]
 209 #define P_MV1 P[9]
 210 #define FLAG_QPEL   1 //must be 1
 211
 212 static int encode_q_branch(SnowContext *s, int level, int x, int y){
 213     uint8_t p_buffer[1024];
 214     uint8_t i_buffer[1024];
 215     uint8_t p_state[sizeof(s->block_state)];
 216     uint8_t i_state[sizeof(s->block_state)];
 217     RangeCoder pc, ic;
 218     uint8_t *pbbak= s->c.bytestream;
 219     uint8_t *pbbak_start= s->c.bytestream_start;
 220     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
 221     const int w= s->b_width  << s->block_max_depth;
 222     const int h= s->b_height << s->block_max_depth;
 223     const int rem_depth= s->block_max_depth - level;
 224     const int index= (x + y*w) << rem_depth;
 225     const int block_w= 1<<(LOG2_MB_SIZE - level);
 226     int trx= (x+1)<<rem_depth;
 227     int try= (y+1)<<rem_depth;
 228     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 229     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 230     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
 231     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
 232     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 233     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 234     int pl = left->color[0];
 235     int pcb= left->color[1];
 236     int pcr= left->color[2];
 237     int pmx, pmy;
 238     int mx=0, my=0;
 239     int l,cr,cb;
 240     const int stride= s->current_picture->linesize[0];
 241     const int uvstride= s->current_picture->linesize[1];
 242     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
 243                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
 244                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
 245     int P[10][2];
 246     int16_t last_mv[3][2];
 247     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
 248     const int shift= 1+qpel;
 249     MotionEstContext *c= &s->m.me;
 250     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 251     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 252     int my_context= av_log2(2*FFABS(left->my - top->my));
 253     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 254     int ref, best_ref, ref_score, ref_mx, ref_my;
 255
 256     av_assert0(sizeof(s->block_state) >= 256);
 257     if(s->keyframe){
 258         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 259         return 0;
 260     }
 261
 262 //    clip predictors / edge ?
 263
 264     P_LEFT[0]= left->mx;
 265     P_LEFT[1]= left->my;
 266     P_TOP [0]= top->mx;
 267     P_TOP [1]= top->my;
 268     P_TOPRIGHT[0]= tr->mx;
 269     P_TOPRIGHT[1]= tr->my;
 270
 271     last_mv[0][0]= s->block[index].mx;
 272     last_mv[0][1]= s->block[index].my;
 273     last_mv[1][0]= right->mx;
 274     last_mv[1][1]= right->my;
 275     last_mv[2][0]= bottom->mx;
 276     last_mv[2][1]= bottom->my;
 277
 278     s->m.mb_stride=2;
 279     s->m.mb_x=
 280     s->m.mb_y= 0;
 281     c->skip= 0;
 282
 283     av_assert1(c->  stride ==   stride);
 284     av_assert1(c->uvstride == uvstride);
 285
 286     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
 287     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
 288     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
 289     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
 290
 291     c->xmin = - x*block_w - 16+3;
 292     c->ymin = - y*block_w - 16+3;
 293     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 294     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 295
 296     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
 297     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
 298     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
 299     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
 300     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
 301     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
 302     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
 303
 304     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
 305     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 306
 307     if (!y) {
 308         c->pred_x= P_LEFT[0];
 309         c->pred_y= P_LEFT[1];
 310     } else {
 311         c->pred_x = P_MEDIAN[0];
 312         c->pred_y = P_MEDIAN[1];
 313     }
 314
 315     score= INT_MAX;
 316     best_ref= 0;
 317     for(ref=0; ref<s->ref_frames; ref++){
 318         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
 319
 320         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
 321                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
 322
 323         av_assert2(ref_mx >= c->xmin);
 324         av_assert2(ref_mx <= c->xmax);
 325         av_assert2(ref_my >= c->ymin);
 326         av_assert2(ref_my <= c->ymax);
 327
 328         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
 329         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
 330         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
 331         if(s->ref_mvs[ref]){
 332             s->ref_mvs[ref][index][0]= ref_mx;
 333             s->ref_mvs[ref][index][1]= ref_my;
 334             s->ref_scores[ref][index]= ref_score;
 335         }
 336         if(score > ref_score){
 337             score= ref_score;
 338             best_ref= ref;
 339             mx= ref_mx;
 340             my= ref_my;
 341         }
 342     }
 343     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
 344
 345   //  subpel search
 346     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
 347     pc= s->c;
 348     pc.bytestream_start=
 349     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
 350     memcpy(p_state, s->block_state, sizeof(s->block_state));
 351
 352     if(level!=s->block_max_depth)
 353         put_rac(&pc, &p_state[4 + s_context], 1);
 354     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
 355     if(s->ref_frames > 1)
 356         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
 357     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
 358     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
 359     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
 360     p_len= pc.bytestream - pc.bytestream_start;
 361     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
 362
 363     block_s= block_w*block_w;
 364     sum = pix_sum(current_data[0], stride, block_w, block_w);
 365     l= (sum + block_s/2)/block_s;
 366     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
 367
 368     if (s->nb_planes > 2) {
 369         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
 370         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 371         cb= (sum + block_s/2)/block_s;
 372     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
 373         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 374         cr= (sum + block_s/2)/block_s;
 375     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
 376     }else
 377         cb = cr = 0;
 378
 379     ic= s->c;
 380     ic.bytestream_start=
 381     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
 382     memcpy(i_state, s->block_state, sizeof(s->block_state));
 383     if(level!=s->block_max_depth)
 384         put_rac(&ic, &i_state[4 + s_context], 1);
 385     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
 386     put_symbol(&ic, &i_state[32],  l-pl , 1);
 387     if (s->nb_planes > 2) {
 388         put_symbol(&ic, &i_state[64], cb-pcb, 1);
 389         put_symbol(&ic, &i_state[96], cr-pcr, 1);
 390     }
 391     i_len= ic.bytestream - ic.bytestream_start;
 392     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
 393
 394 //    assert(score==256*256*256*64-1);
 395     av_assert1(iscore < 255*255*256 + s->lambda2*10);
 396     av_assert1(iscore >= 0);
 397     av_assert1(l>=0 && l<=255);
 398     av_assert1(pl>=0 && pl<=255);
 399
 400     if(level==0){
 401         int varc= iscore >> 8;
 402         int vard= score >> 8;
 403         if (vard <= 64 || vard < varc)
 404             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
 405         else
 406             c->scene_change_score+= s->m.qscale;
 407     }
 408
 409     if(level!=s->block_max_depth){
 410         put_rac(&s->c, &s->block_state[4 + s_context], 0);
 411         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
 412         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
 413         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
 414         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
 415         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
 416
 417         if(score2 < score && score2 < iscore)
 418             return score2;
 419     }
 420
 421     if(iscore < score){
 422         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 423         memcpy(pbbak, i_buffer, i_len);
 424         s->c= ic;
 425         s->c.bytestream_start= pbbak_start;
 426         s->c.bytestream= pbbak + i_len;
 427         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
 428         memcpy(s->block_state, i_state, sizeof(s->block_state));
 429         return iscore;
 430     }else{
 431         memcpy(pbbak, p_buffer, p_len);
 432         s->c= pc;
 433         s->c.bytestream_start= pbbak_start;
 434         s->c.bytestream= pbbak + p_len;
 435         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
 436         memcpy(s->block_state, p_state, sizeof(s->block_state));
 437         return score;
 438     }
 439 }
 440
 441 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
 442     const int w= s->b_width  << s->block_max_depth;
 443     const int rem_depth= s->block_max_depth - level;
 444     const int index= (x + y*w) << rem_depth;
 445     int trx= (x+1)<<rem_depth;
 446     BlockNode *b= &s->block[index];
 447     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 448     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 449     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 450     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 451     int pl = left->color[0];
 452     int pcb= left->color[1];
 453     int pcr= left->color[2];
 454     int pmx, pmy;
 455     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 456     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
 457     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
 458     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 459
 460     if(s->keyframe){
 461         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 462         return;
 463     }
 464
 465     if(level!=s->block_max_depth){
 466         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
 467             put_rac(&s->c, &s->block_state[4 + s_context], 1);
 468         }else{
 469             put_rac(&s->c, &s->block_state[4 + s_context], 0);
 470             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
 471             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
 472             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
 473             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
 474             return;
 475         }
 476     }
 477     if(b->type & BLOCK_INTRA){
 478         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 479         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
 480         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
 481         if (s->nb_planes > 2) {
 482             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
 483             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
 484         }
 485         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
 486     }else{
 487         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
 488         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
 489         if(s->ref_frames > 1)
 490             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
 491         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
 492         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
 493         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
 494     }
 495 }
 496
 497 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
 498     int i, x2, y2;
 499     Plane *p= &s->plane[plane_index];
 500     const int block_size = MB_SIZE >> s->block_max_depth;
 501     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 502     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 503     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 504     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 505     const int ref_stride= s->current_picture->linesize[plane_index];
 506     uint8_t *src= s-> input_picture->data[plane_index];
 507     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
 508     const int b_stride = s->b_width << s->block_max_depth;
 509     const int w= p->width;
 510     const int h= p->height;
 511     int index= mb_x + mb_y*b_stride;
 512     BlockNode *b= &s->block[index];
 513     BlockNode backup= *b;
 514     int ab=0;
 515     int aa=0;
 516
 517     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
 518
 519     b->type|= BLOCK_INTRA;
 520     b->color[plane_index]= 0;
 521     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
 522
 523     for(i=0; i<4; i++){
 524         int mb_x2= mb_x + (i &1) - 1;
 525         int mb_y2= mb_y + (i>>1) - 1;
 526         int x= block_w*mb_x2 + block_w/2;
 527         int y= block_h*mb_y2 + block_h/2;
 528
 529         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
 530                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
 531
 532         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
 533             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
 534                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
 535                 int obmc_v= obmc[index];
 536                 int d;
 537                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
 538                 if(x<0) obmc_v += obmc[index + block_w];
 539                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
 540                 if(x+block_w>w) obmc_v += obmc[index - block_w];
 541                 //FIXME precalculate this or simplify it somehow else
 542
 543                 d = -dst[index] + (1<<(FRAC_BITS-1));
 544                 dst[index] = d;
 545                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
 546                 aa += obmc_v * obmc_v; //FIXME precalculate this
 547             }
 548         }
 549     }
 550     *b= backup;
 551
 552     return av_clip( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa), 0, 255); //FIXME we should not need clipping
 553 }
 554
 555 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
 556     const int b_stride = s->b_width << s->block_max_depth;
 557     const int b_height = s->b_height<< s->block_max_depth;
 558     int index= x + y*b_stride;
 559     const BlockNode *b     = &s->block[index];
 560     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 561     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
 562     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
 563     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
 564     int dmx, dmy;
 565 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 566 //  int my_context= av_log2(2*FFABS(left->my - top->my));
 567
 568     if(x<0 || x>=b_stride || y>=b_height)
 569         return 0;
 570 /*
 571 1            0      0
 572 01X          1-2    1
 573 001XX        3-6    2-3
 574 0001XXX      7-14   4-7
 575 00001XXXX   15-30   8-15
 576 */
 577 //FIXME try accurate rate
 578 //FIXME intra and inter predictors if surrounding blocks are not the same type
 579     if(b->type & BLOCK_INTRA){
 580         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
 581                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
 582                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
 583     }else{
 584         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
 585         dmx-= b->mx;
 586         dmy-= b->my;
 587         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
 588                     + av_log2(2*FFABS(dmy))
 589                     + av_log2(2*b->ref));
 590     }
 591 }
 592
 593 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
 594     Plane *p= &s->plane[plane_index];
 595     const int block_size = MB_SIZE >> s->block_max_depth;
 596     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 597     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 598     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 599     const int ref_stride= s->current_picture->linesize[plane_index];
 600     uint8_t *dst= s->current_picture->data[plane_index];
 601     uint8_t *src= s->  input_picture->data[plane_index];
 602     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
 603     uint8_t *cur = s->scratchbuf;
 604     uint8_t *tmp = s->emu_edge_buffer;
 605     const int b_stride = s->b_width << s->block_max_depth;
 606     const int b_height = s->b_height<< s->block_max_depth;
 607     const int w= p->width;
 608     const int h= p->height;
 609     int distortion;
 610     int rate= 0;
 611     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 612     int sx= block_w*mb_x - block_w/2;
 613     int sy= block_h*mb_y - block_h/2;
 614     int x0= FFMAX(0,-sx);
 615     int y0= FFMAX(0,-sy);
 616     int x1= FFMIN(block_w*2, w-sx);
 617     int y1= FFMIN(block_h*2, h-sy);
 618     int i,x,y;
 619
 620     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
 621
 622     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);
 623
 624     for(y=y0; y<y1; y++){
 625         const uint8_t *obmc1= obmc_edged[y];
 626         const IDWTELEM *pred1 = pred + y*obmc_stride;
 627         uint8_t *cur1 = cur + y*ref_stride;
 628         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
 629         for(x=x0; x<x1; x++){
 630 #if FRAC_BITS >= LOG2_OBMC_MAX
 631             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
 632 #else
 633             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
 634 #endif
 635             v = (v + pred1[x]) >> FRAC_BITS;
 636             if(v&(~255)) v= ~(v>>31);
 637             dst1[x] = v;
 638         }
 639     }
 640
 641     /* copy the regions where obmc[] = (uint8_t)256 */
 642     if(LOG2_OBMC_MAX == 8
 643         && (mb_x == 0 || mb_x == b_stride-1)
 644         && (mb_y == 0 || mb_y == b_height-1)){
 645         if(mb_x == 0)
 646             x1 = block_w;
 647         else
 648             x0 = block_w;
 649         if(mb_y == 0)
 650             y1 = block_h;
 651         else
 652             y0 = block_h;
 653         for(y=y0; y<y1; y++)
 654             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
 655     }
 656
 657     if(block_w==16){
 658         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
 659         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
 660         /* FIXME cmps overlap but do not cover the wavelet's whole support.
 661          * So improving the score of one block is not strictly guaranteed
 662          * to improve the score of the whole frame, thus iterative motion
 663          * estimation does not always converge. */
 664         if(s->avctx->me_cmp == FF_CMP_W97)
 665             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 666         else if(s->avctx->me_cmp == FF_CMP_W53)
 667             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 668         else{
 669             distortion = 0;
 670             for(i=0; i<4; i++){
 671                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
 672                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
 673             }
 674         }
 675     }else{
 676         av_assert2(block_w==8);
 677         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
 678     }
 679
 680     if(plane_index==0){
 681         for(i=0; i<4; i++){
 682 /* ..RRr
 683  * .RXx.
 684  * rxx..
 685  */
 686             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
 687         }
 688         if(mb_x == b_stride-2)
 689             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
 690     }
 691     return distortion + rate*penalty_factor;
 692 }
 693
 694 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
 695     int i, y2;
 696     Plane *p= &s->plane[plane_index];
 697     const int block_size = MB_SIZE >> s->block_max_depth;
 698     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 699     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 700     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 701     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 702     const int ref_stride= s->current_picture->linesize[plane_index];
 703     uint8_t *dst= s->current_picture->data[plane_index];
 704     uint8_t *src= s-> input_picture->data[plane_index];
 705     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
 706     // const has only been removed from zero_dst to suppress a warning
 707     static IDWTELEM zero_dst[4096]; //FIXME
 708     const int b_stride = s->b_width << s->block_max_depth;
 709     const int w= p->width;
 710     const int h= p->height;
 711     int distortion= 0;
 712     int rate= 0;
 713     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 714
 715     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
 716
 717     for(i=0; i<9; i++){
 718         int mb_x2= mb_x + (i%3) - 1;
 719         int mb_y2= mb_y + (i/3) - 1;
 720         int x= block_w*mb_x2 + block_w/2;
 721         int y= block_h*mb_y2 + block_h/2;
 722
 723         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
 724                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
 725
 726         //FIXME find a cleaner/simpler way to skip the outside stuff
 727         for(y2= y; y2<0; y2++)
 728             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 729         for(y2= h; y2<y+block_h; y2++)
 730             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 731         if(x<0){
 732             for(y2= y; y2<y+block_h; y2++)
 733                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
 734         }
 735         if(x+block_w > w){
 736             for(y2= y; y2<y+block_h; y2++)
 737                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
 738         }
 739
 740         av_assert1(block_w== 8 || block_w==16);
 741         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
 742     }
 743
 744     if(plane_index==0){
 745         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
 746         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
 747
 748 /* ..RRRr
 749  * .RXXx.
 750  * .RXXx.
 751  * rxxx.
 752  */
 753         if(merged)
 754             rate = get_block_bits(s, mb_x, mb_y, 2);
 755         for(i=merged?4:0; i<9; i++){
 756             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
 757             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
 758         }
 759     }
 760     return distortion + rate*penalty_factor;
 761 }
 762
 763 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 764     const int w= b->width;
 765     const int h= b->height;
 766     int x, y;
 767
 768     if(1){
 769         int run=0;
 770         int *runs = s->run_buffer;
 771         int run_index=0;
 772         int max_index;
 773
 774         for(y=0; y<h; y++){
 775             for(x=0; x<w; x++){
 776                 int v, p=0;
 777                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 778                 v= src[x + y*stride];
 779
 780                 if(y){
 781                     t= src[x + (y-1)*stride];
 782                     if(x){
 783                         lt= src[x - 1 + (y-1)*stride];
 784                     }
 785                     if(x + 1 < w){
 786                         rt= src[x + 1 + (y-1)*stride];
 787                     }
 788                 }
 789                 if(x){
 790                     l= src[x - 1 + y*stride];
 791                     /*if(x > 1){
 792                         if(orientation==1) ll= src[y + (x-2)*stride];
 793                         else               ll= src[x - 2 + y*stride];
 794                     }*/
 795                 }
 796                 if(parent){
 797                     int px= x>>1;
 798                     int py= y>>1;
 799                     if(px<b->parent->width && py<b->parent->height)
 800                         p= parent[px + py*2*stride];
 801                 }
 802                 if(!(/*ll|*/l|lt|t|rt|p)){
 803                     if(v){
 804                         runs[run_index++]= run;
 805                         run=0;
 806                     }else{
 807                         run++;
 808                     }
 809                 }
 810             }
 811         }
 812         max_index= run_index;
 813         runs[run_index++]= run;
 814         run_index=0;
 815         run= runs[run_index++];
 816
 817         put_symbol2(&s->c, b->state[30], max_index, 0);
 818         if(run_index <= max_index)
 819             put_symbol2(&s->c, b->state[1], run, 3);
 820
 821         for(y=0; y<h; y++){
 822             if(s->c.bytestream_end - s->c.bytestream < w*40){
 823                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 824                 return -1;
 825             }
 826             for(x=0; x<w; x++){
 827                 int v, p=0;
 828                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 829                 v= src[x + y*stride];
 830
 831                 if(y){
 832                     t= src[x + (y-1)*stride];
 833                     if(x){
 834                         lt= src[x - 1 + (y-1)*stride];
 835                     }
 836                     if(x + 1 < w){
 837                         rt= src[x + 1 + (y-1)*stride];
 838                     }
 839                 }
 840                 if(x){
 841                     l= src[x - 1 + y*stride];
 842                     /*if(x > 1){
 843                         if(orientation==1) ll= src[y + (x-2)*stride];
 844                         else               ll= src[x - 2 + y*stride];
 845                     }*/
 846                 }
 847                 if(parent){
 848                     int px= x>>1;
 849                     int py= y>>1;
 850                     if(px<b->parent->width && py<b->parent->height)
 851                         p= parent[px + py*2*stride];
 852                 }
 853                 if(/*ll|*/l|lt|t|rt|p){
 854                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 855
 856                     put_rac(&s->c, &b->state[0][context], !!v);
 857                 }else{
 858                     if(!run){
 859                         run= runs[run_index++];
 860
 861                         if(run_index <= max_index)
 862                             put_symbol2(&s->c, b->state[1], run, 3);
 863                         av_assert2(v);
 864                     }else{
 865                         run--;
 866                         av_assert2(!v);
 867                     }
 868                 }
 869                 if(v){
 870                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 871                     int l2= 2*FFABS(l) + (l<0);
 872                     int t2= 2*FFABS(t) + (t<0);
 873
 874                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
 875                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
 876                 }
 877             }
 878         }
 879     }
 880     return 0;
 881 }
 882
 883 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 884 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
 885 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
 886     return encode_subband_c0run(s, b, src, parent, stride, orientation);
 887 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
 888 }
 889
 890 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){
 891     const int b_stride= s->b_width << s->block_max_depth;
 892     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 893     BlockNode backup= *block;
 894     unsigned value;
 895     int rd, index;
 896
 897     av_assert2(mb_x>=0 && mb_y>=0);
 898     av_assert2(mb_x<b_stride);
 899
 900     if(intra){
 901         block->color[0] = p[0];
 902         block->color[1] = p[1];
 903         block->color[2] = p[2];
 904         block->type |= BLOCK_INTRA;
 905     }else{
 906         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
 907         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
 908         if(s->me_cache[index] == value)
 909             return 0;
 910         s->me_cache[index]= value;
 911
 912         block->mx= p[0];
 913         block->my= p[1];
 914         block->type &= ~BLOCK_INTRA;
 915     }
 916
 917     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
 918
 919 //FIXME chroma
 920     if(rd < *best_rd){
 921         *best_rd= rd;
 922         return 1;
 923     }else{
 924         *block= backup;
 925         return 0;
 926     }
 927 }
 928
 929 /* special case for int[2] args we discard afterwards,
 930  * fixes compilation problem with gcc 2.95 */
 931 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){
 932     int p[2] = {p0, p1};
 933     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
 934 }
 935
 936 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){
 937     const int b_stride= s->b_width << s->block_max_depth;
 938     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 939     BlockNode backup[4];
 940     unsigned value;
 941     int rd, index;
 942
 943     /* We don't initialize backup[] during variable declaration, because
 944      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
 945      * 'int16_t'". */
 946     backup[0] = block[0];
 947     backup[1] = block[1];
 948     backup[2] = block[b_stride];
 949     backup[3] = block[b_stride + 1];
 950
 951     av_assert2(mb_x>=0 && mb_y>=0);
 952     av_assert2(mb_x<b_stride);
 953     av_assert2(((mb_x|mb_y)&1) == 0);
 954
 955     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
 956     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
 957     if(s->me_cache[index] == value)
 958         return 0;
 959     s->me_cache[index]= value;
 960
 961     block->mx= p0;
 962     block->my= p1;
 963     block->ref= ref;
 964     block->type &= ~BLOCK_INTRA;
 965     block[1]= block[b_stride]= block[b_stride+1]= *block;
 966
 967     rd= get_4block_rd(s, mb_x, mb_y, 0);
 968
 969 //FIXME chroma
 970     if(rd < *best_rd){
 971         *best_rd= rd;
 972         return 1;
 973     }else{
 974         block[0]= backup[0];
 975         block[1]= backup[1];
 976         block[b_stride]= backup[2];
 977         block[b_stride+1]= backup[3];
 978         return 0;
 979     }
 980 }
 981
 982 static void iterative_me(SnowContext *s){
 983     int pass, mb_x, mb_y;
 984     const int b_width = s->b_width  << s->block_max_depth;
 985     const int b_height= s->b_height << s->block_max_depth;
 986     const int b_stride= b_width;
 987     int color[3];
 988
 989     {
 990         RangeCoder r = s->c;
 991         uint8_t state[sizeof(s->block_state)];
 992         memcpy(state, s->block_state, sizeof(s->block_state));
 993         for(mb_y= 0; mb_y<s->b_height; mb_y++)
 994             for(mb_x= 0; mb_x<s->b_width; mb_x++)
 995                 encode_q_branch(s, 0, mb_x, mb_y);
 996         s->c = r;
 997         memcpy(s->block_state, state, sizeof(s->block_state));
 998     }
 999
1000     for(pass=0; pass<25; pass++){
1001         int change= 0;
1002
1003         for(mb_y= 0; mb_y<b_height; mb_y++){
1004             for(mb_x= 0; mb_x<b_width; mb_x++){
1005                 int dia_change, i, j, ref;
1006                 int best_rd= INT_MAX, ref_rd;
1007                 BlockNode backup, ref_b;
1008                 const int index= mb_x + mb_y * b_stride;
1009                 BlockNode *block= &s->block[index];
1010                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1011                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1012                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1013                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1014                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1015                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1016                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1017                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1018                 const int b_w= (MB_SIZE >> s->block_max_depth);
1019                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1020
1021                 if(pass && (block->type & BLOCK_OPT))
1022                     continue;
1023                 block->type |= BLOCK_OPT;
1024
1025                 backup= *block;
1026
1027                 if(!s->me_cache_generation)
1028                     memset(s->me_cache, 0, sizeof(s->me_cache));
1029                 s->me_cache_generation += 1<<22;
1030
1031                 //FIXME precalculate
1032                 {
1033                     int x, y;
1034                     for (y = 0; y < b_w * 2; y++)
1035                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1036                     if(mb_x==0)
1037                         for(y=0; y<b_w*2; y++)
1038                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1039                     if(mb_x==b_stride-1)
1040                         for(y=0; y<b_w*2; y++)
1041                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1042                     if(mb_y==0){
1043                         for(x=0; x<b_w*2; x++)
1044                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1045                         for(y=1; y<b_w; y++)
1046                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1047                     }
1048                     if(mb_y==b_height-1){
1049                         for(x=0; x<b_w*2; x++)
1050                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1051                         for(y=b_w; y<b_w*2-1; y++)
1052                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1053                     }
1054                 }
1055
1056                 //skip stuff outside the picture
1057                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1058                     uint8_t *src= s->  input_picture->data[0];
1059                     uint8_t *dst= s->current_picture->data[0];
1060                     const int stride= s->current_picture->linesize[0];
1061                     const int block_w= MB_SIZE >> s->block_max_depth;
1062                     const int block_h= MB_SIZE >> s->block_max_depth;
1063                     const int sx= block_w*mb_x - block_w/2;
1064                     const int sy= block_h*mb_y - block_h/2;
1065                     const int w= s->plane[0].width;
1066                     const int h= s->plane[0].height;
1067                     int y;
1068
1069                     for(y=sy; y<0; y++)
1070                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1071                     for(y=h; y<sy+block_h*2; y++)
1072                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1073                     if(sx<0){
1074                         for(y=sy; y<sy+block_h*2; y++)
1075                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1076                     }
1077                     if(sx+block_w*2 > w){
1078                         for(y=sy; y<sy+block_h*2; y++)
1079                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1080                     }
1081                 }
1082
1083                 // intra(black) = neighbors' contribution to the current block
1084                 for(i=0; i < s->nb_planes; i++)
1085                     color[i]= get_dc(s, mb_x, mb_y, i);
1086
1087                 // get previous score (cannot be cached due to OBMC)
1088                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1089                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1090                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1091                 }else
1092                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1093
1094                 ref_b= *block;
1095                 ref_rd= best_rd;
1096                 for(ref=0; ref < s->ref_frames; ref++){
1097                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1098                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1099                         continue;
1100                     block->ref= ref;
1101                     best_rd= INT_MAX;
1102
1103                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1104                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1105                     if(tb)
1106                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1107                     if(lb)
1108                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1109                     if(rb)
1110                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1111                     if(bb)
1112                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1113
1114                     /* fullpel ME */
1115                     //FIXME avoid subpel interpolation / round to nearest integer
1116                     do{
1117                         dia_change=0;
1118                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1119                             for(j=0; j<i; j++){
1120                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1121                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1122                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1123                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1124                             }
1125                         }
1126                     }while(dia_change);
1127                     /* subpel ME */
1128                     do{
1129                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1130                         dia_change=0;
1131                         for(i=0; i<8; i++)
1132                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1133                     }while(dia_change);
1134                     //FIXME or try the standard 2 pass qpel or similar
1135
1136                     mvr[0][0]= block->mx;
1137                     mvr[0][1]= block->my;
1138                     if(ref_rd > best_rd){
1139                         ref_rd= best_rd;
1140                         ref_b= *block;
1141                     }
1142                 }
1143                 best_rd= ref_rd;
1144                 *block= ref_b;
1145                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1146                 //FIXME RD style color selection
1147                 if(!same_block(block, &backup)){
1148                     if(tb ) tb ->type &= ~BLOCK_OPT;
1149                     if(lb ) lb ->type &= ~BLOCK_OPT;
1150                     if(rb ) rb ->type &= ~BLOCK_OPT;
1151                     if(bb ) bb ->type &= ~BLOCK_OPT;
1152                     if(tlb) tlb->type &= ~BLOCK_OPT;
1153                     if(trb) trb->type &= ~BLOCK_OPT;
1154                     if(blb) blb->type &= ~BLOCK_OPT;
1155                     if(brb) brb->type &= ~BLOCK_OPT;
1156                     change ++;
1157                 }
1158             }
1159         }
1160         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1161         if(!change)
1162             break;
1163     }
1164
1165     if(s->block_max_depth == 1){
1166         int change= 0;
1167         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1168             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1169                 int i;
1170                 int best_rd, init_rd;
1171                 const int index= mb_x + mb_y * b_stride;
1172                 BlockNode *b[4];
1173
1174                 b[0]= &s->block[index];
1175                 b[1]= b[0]+1;
1176                 b[2]= b[0]+b_stride;
1177                 b[3]= b[2]+1;
1178                 if(same_block(b[0], b[1]) &&
1179                    same_block(b[0], b[2]) &&
1180                    same_block(b[0], b[3]))
1181                     continue;
1182
1183                 if(!s->me_cache_generation)
1184                     memset(s->me_cache, 0, sizeof(s->me_cache));
1185                 s->me_cache_generation += 1<<22;
1186
1187                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1188
1189                 //FIXME more multiref search?
1190                 check_4block_inter(s, mb_x, mb_y,
1191                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1192                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1193
1194                 for(i=0; i<4; i++)
1195                     if(!(b[i]->type&BLOCK_INTRA))
1196                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1197
1198                 if(init_rd != best_rd)
1199                     change++;
1200             }
1201         }
1202         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1203     }
1204 }
1205
1206 static void encode_blocks(SnowContext *s, int search){
1207     int x, y;
1208     int w= s->b_width;
1209     int h= s->b_height;
1210
1211     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1212         iterative_me(s);
1213
1214     for(y=0; y<h; y++){
1215         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1216             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1217             return;
1218         }
1219         for(x=0; x<w; x++){
1220             if(s->avctx->me_method == ME_ITER || !search)
1221                 encode_q_branch2(s, 0, x, y);
1222             else
1223                 encode_q_branch (s, 0, x, y);
1224         }
1225     }
1226 }
1227
1228 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1229     const int w= b->width;
1230     const int h= b->height;
1231     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1232     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1233     int x,y, thres1, thres2;
1234
1235     if(s->qlog == LOSSLESS_QLOG){
1236         for(y=0; y<h; y++)
1237             for(x=0; x<w; x++)
1238                 dst[x + y*stride]= src[x + y*stride];
1239         return;
1240     }
1241
1242     bias= bias ? 0 : (3*qmul)>>3;
1243     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1244     thres2= 2*thres1;
1245
1246     if(!bias){
1247         for(y=0; y<h; y++){
1248             for(x=0; x<w; x++){
1249                 int i= src[x + y*stride];
1250
1251                 if((unsigned)(i+thres1) > thres2){
1252                     if(i>=0){
1253                         i<<= QEXPSHIFT;
1254                         i/= qmul; //FIXME optimize
1255                         dst[x + y*stride]=  i;
1256                     }else{
1257                         i= -i;
1258                         i<<= QEXPSHIFT;
1259                         i/= qmul; //FIXME optimize
1260                         dst[x + y*stride]= -i;
1261                     }
1262                 }else
1263                     dst[x + y*stride]= 0;
1264             }
1265         }
1266     }else{
1267         for(y=0; y<h; y++){
1268             for(x=0; x<w; x++){
1269                 int i= src[x + y*stride];
1270
1271                 if((unsigned)(i+thres1) > thres2){
1272                     if(i>=0){
1273                         i<<= QEXPSHIFT;
1274                         i= (i + bias) / qmul; //FIXME optimize
1275                         dst[x + y*stride]=  i;
1276                     }else{
1277                         i= -i;
1278                         i<<= QEXPSHIFT;
1279                         i= (i + bias) / qmul; //FIXME optimize
1280                         dst[x + y*stride]= -i;
1281                     }
1282                 }else
1283                     dst[x + y*stride]= 0;
1284             }
1285         }
1286     }
1287 }
1288
1289 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1290     const int w= b->width;
1291     const int h= b->height;
1292     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1293     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1294     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1295     int x,y;
1296
1297     if(s->qlog == LOSSLESS_QLOG) return;
1298
1299     for(y=0; y<h; y++){
1300         for(x=0; x<w; x++){
1301             int i= src[x + y*stride];
1302             if(i<0){
1303                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1304             }else if(i>0){
1305                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1306             }
1307         }
1308     }
1309 }
1310
1311 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1312     const int w= b->width;
1313     const int h= b->height;
1314     int x,y;
1315
1316     for(y=h-1; y>=0; y--){
1317         for(x=w-1; x>=0; x--){
1318             int i= x + y*stride;
1319
1320             if(x){
1321                 if(use_median){
1322                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1323                     else  src[i] -= src[i - 1];
1324                 }else{
1325                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1326                     else  src[i] -= src[i - 1];
1327                 }
1328             }else{
1329                 if(y) src[i] -= src[i - stride];
1330             }
1331         }
1332     }
1333 }
1334
1335 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1336     const int w= b->width;
1337     const int h= b->height;
1338     int x,y;
1339
1340     for(y=0; y<h; y++){
1341         for(x=0; x<w; x++){
1342             int i= x + y*stride;
1343
1344             if(x){
1345                 if(use_median){
1346                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1347                     else  src[i] += src[i - 1];
1348                 }else{
1349                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1350                     else  src[i] += src[i - 1];
1351                 }
1352             }else{
1353                 if(y) src[i] += src[i - stride];
1354             }
1355         }
1356     }
1357 }
1358
1359 static void encode_qlogs(SnowContext *s){
1360     int plane_index, level, orientation;
1361
1362     for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1363         for(level=0; level<s->spatial_decomposition_count; level++){
1364             for(orientation=level ? 1:0; orientation<4; orientation++){
1365                 if(orientation==2) continue;
1366                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1367             }
1368         }
1369     }
1370 }
1371
1372 static void encode_header(SnowContext *s){
1373     int plane_index, i;
1374     uint8_t kstate[32];
1375
1376     memset(kstate, MID_STATE, sizeof(kstate));
1377
1378     put_rac(&s->c, kstate, s->keyframe);
1379     if(s->keyframe || s->always_reset){
1380         ff_snow_reset_contexts(s);
1381         s->last_spatial_decomposition_type=
1382         s->last_qlog=
1383         s->last_qbias=
1384         s->last_mv_scale=
1385         s->last_block_max_depth= 0;
1386         for(plane_index=0; plane_index<2; plane_index++){
1387             Plane *p= &s->plane[plane_index];
1388             p->last_htaps=0;
1389             p->last_diag_mc=0;
1390             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1391         }
1392     }
1393     if(s->keyframe){
1394         put_symbol(&s->c, s->header_state, s->version, 0);
1395         put_rac(&s->c, s->header_state, s->always_reset);
1396         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1397         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1398         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1399         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1400         if (s->nb_planes > 2) {
1401             put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1402             put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1403         }
1404         put_rac(&s->c, s->header_state, s->spatial_scalability);
1405 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1406         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1407
1408         encode_qlogs(s);
1409     }
1410
1411     if(!s->keyframe){
1412         int update_mc=0;
1413         for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1414             Plane *p= &s->plane[plane_index];
1415             update_mc |= p->last_htaps   != p->htaps;
1416             update_mc |= p->last_diag_mc != p->diag_mc;
1417             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1418         }
1419         put_rac(&s->c, s->header_state, update_mc);
1420         if(update_mc){
1421             for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1422                 Plane *p= &s->plane[plane_index];
1423                 put_rac(&s->c, s->header_state, p->diag_mc);
1424                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1425                 for(i= p->htaps/2; i; i--)
1426                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1427             }
1428         }
1429         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1430             put_rac(&s->c, s->header_state, 1);
1431             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1432             encode_qlogs(s);
1433         }else
1434             put_rac(&s->c, s->header_state, 0);
1435     }
1436
1437     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1438     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1439     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1440     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1441     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1442
1443 }
1444
1445 static void update_last_header_values(SnowContext *s){
1446     int plane_index;
1447
1448     if(!s->keyframe){
1449         for(plane_index=0; plane_index<2; plane_index++){
1450             Plane *p= &s->plane[plane_index];
1451             p->last_diag_mc= p->diag_mc;
1452             p->last_htaps  = p->htaps;
1453             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1454         }
1455     }
1456
1457     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1458     s->last_qlog                        = s->qlog;
1459     s->last_qbias                       = s->qbias;
1460     s->last_mv_scale                    = s->mv_scale;
1461     s->last_block_max_depth             = s->block_max_depth;
1462     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1463 }
1464
1465 static int qscale2qlog(int qscale){
1466     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1467            + 61*QROOT/8; ///< 64 > 60
1468 }
1469
1470 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1471 {
1472     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1473      * FIXME we know exact mv bits at this point,
1474      * but ratecontrol isn't set up to include them. */
1475     uint32_t coef_sum= 0;
1476     int level, orientation, delta_qlog;
1477
1478     for(level=0; level<s->spatial_decomposition_count; level++){
1479         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1480             SubBand *b= &s->plane[0].band[level][orientation];
1481             IDWTELEM *buf= b->ibuf;
1482             const int w= b->width;
1483             const int h= b->height;
1484             const int stride= b->stride;
1485             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1486             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1487             const int qdiv= (1<<16)/qmul;
1488             int x, y;
1489             //FIXME this is ugly
1490             for(y=0; y<h; y++)
1491                 for(x=0; x<w; x++)
1492                     buf[x+y*stride]= b->buf[x+y*stride];
1493             if(orientation==0)
1494                 decorrelate(s, b, buf, stride, 1, 0);
1495             for(y=0; y<h; y++)
1496                 for(x=0; x<w; x++)
1497                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1498         }
1499     }
1500
1501     /* ugly, ratecontrol just takes a sqrt again */
1502     av_assert0(coef_sum < INT_MAX);
1503     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1504
1505     if(pict->pict_type == AV_PICTURE_TYPE_I){
1506         s->m.current_picture.mb_var_sum= coef_sum;
1507         s->m.current_picture.mc_mb_var_sum= 0;
1508     }else{
1509         s->m.current_picture.mc_mb_var_sum= coef_sum;
1510         s->m.current_picture.mb_var_sum= 0;
1511     }
1512
1513     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1514     if (pict->quality < 0)
1515         return INT_MIN;
1516     s->lambda= pict->quality * 3/2;
1517     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1518     s->qlog+= delta_qlog;
1519     return delta_qlog;
1520 }
1521
1522 static void calculate_visual_weight(SnowContext *s, Plane *p){
1523     int width = p->width;
1524     int height= p->height;
1525     int level, orientation, x, y;
1526
1527     for(level=0; level<s->spatial_decomposition_count; level++){
1528         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1529             SubBand *b= &p->band[level][orientation];
1530             IDWTELEM *ibuf= b->ibuf;
1531             int64_t error=0;
1532
1533             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1534             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1535             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1536             for(y=0; y<height; y++){
1537                 for(x=0; x<width; x++){
1538                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1539                     error += d*d;
1540                 }
1541             }
1542
1543             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1544         }
1545     }
1546 }
1547
1548 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1549                         const AVFrame *pict, int *got_packet)
1550 {
1551     SnowContext *s = avctx->priv_data;
1552     RangeCoder * const c= &s->c;
1553     AVFrame *pic = pict;
1554     const int width= s->avctx->width;
1555     const int height= s->avctx->height;
1556     int level, orientation, plane_index, i, y, ret;
1557     uint8_t rc_header_bak[sizeof(s->header_state)];
1558     uint8_t rc_block_bak[sizeof(s->block_state)];
1559
1560     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1561         return ret;
1562
1563     ff_init_range_encoder(c, pkt->data, pkt->size);
1564     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1565
1566     for(i=0; i < s->nb_planes; i++){
1567         int hshift= i ? s->chroma_h_shift : 0;
1568         int vshift= i ? s->chroma_v_shift : 0;
1569         for(y=0; y<(height>>vshift); y++)
1570             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1571                    &pict->data[i][y * pict->linesize[i]],
1572                    width>>hshift);
1573         s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1574                                 width >> hshift, height >> vshift,
1575                                 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1576                                 EDGE_TOP | EDGE_BOTTOM);
1577
1578     }
1579     emms_c();
1580     s->new_picture = pict;
1581
1582     s->m.picture_number= avctx->frame_number;
1583     if(avctx->flags&CODEC_FLAG_PASS2){
1584         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1585         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1586         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1587             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1588             if (pic->quality < 0)
1589                 return -1;
1590         }
1591     }else{
1592         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1593         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1594     }
1595
1596     if(s->pass1_rc && avctx->frame_number == 0)
1597         pic->quality = 2*FF_QP2LAMBDA;
1598     if (pic->quality) {
1599         s->qlog   = qscale2qlog(pic->quality);
1600         s->lambda = pic->quality * 3/2;
1601     }
1602     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1603         s->qlog= LOSSLESS_QLOG;
1604         s->lambda = 0;
1605     }//else keep previous frame's qlog until after motion estimation
1606
1607     if (s->current_picture->data[0] && !(s->avctx->flags&CODEC_FLAG_EMU_EDGE)) {
1608         int w = s->avctx->width;
1609         int h = s->avctx->height;
1610
1611         s->mpvencdsp.draw_edges(s->current_picture->data[0],
1612                                 s->current_picture->linesize[0], w   , h   ,
1613                                 EDGE_WIDTH  , EDGE_WIDTH  , EDGE_TOP | EDGE_BOTTOM);
1614         if (s->current_picture->data[2]) {
1615             s->mpvencdsp.draw_edges(s->current_picture->data[1],
1616                                     s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1617                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1618             s->mpvencdsp.draw_edges(s->current_picture->data[2],
1619                                     s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1620                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1621         }
1622     }
1623
1624     ff_snow_frame_start(s);
1625     avctx->coded_frame= s->current_picture;
1626
1627     s->m.current_picture_ptr= &s->m.current_picture;
1628     s->m.current_picture.f = s->current_picture;
1629     s->m.current_picture.f->pts = pict->pts;
1630     if(pic->pict_type == AV_PICTURE_TYPE_P){
1631         int block_width = (width +15)>>4;
1632         int block_height= (height+15)>>4;
1633         int stride= s->current_picture->linesize[0];
1634
1635         av_assert0(s->current_picture->data[0]);
1636         av_assert0(s->last_picture[0]->data[0]);
1637
1638         s->m.avctx= s->avctx;
1639         s->m.   last_picture.f = s->last_picture[0];
1640         s->m.    new_picture.f = s->input_picture;
1641         s->m.   last_picture_ptr= &s->m.   last_picture;
1642         s->m.linesize = stride;
1643         s->m.uvlinesize= s->current_picture->linesize[1];
1644         s->m.width = width;
1645         s->m.height= height;
1646         s->m.mb_width = block_width;
1647         s->m.mb_height= block_height;
1648         s->m.mb_stride=   s->m.mb_width+1;
1649         s->m.b8_stride= 2*s->m.mb_width+1;
1650         s->m.f_code=1;
1651         s->m.pict_type = pic->pict_type;
1652         s->m.me_method= s->avctx->me_method;
1653         s->m.me.scene_change_score=0;
1654         s->m.flags= s->avctx->flags;
1655         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1656         s->m.out_format= FMT_H263;
1657         s->m.unrestricted_mv= 1;
1658
1659         s->m.lambda = s->lambda;
1660         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1661         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1662
1663         s->m.dsp= s->dsp; //move
1664         s->m.qdsp= s->qdsp; //move
1665         s->m.hdsp = s->hdsp;
1666         ff_init_me(&s->m);
1667         s->hdsp = s->m.hdsp;
1668         s->dsp= s->m.dsp;
1669     }
1670
1671     if(s->pass1_rc){
1672         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1673         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1674     }
1675
1676 redo_frame:
1677
1678     s->spatial_decomposition_count= 5;
1679
1680     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1681           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1682         s->spatial_decomposition_count--;
1683
1684     if (s->spatial_decomposition_count <= 0) {
1685         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1686         return AVERROR(EINVAL);
1687     }
1688
1689     s->m.pict_type = pic->pict_type;
1690     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1691
1692     ff_snow_common_init_after_header(avctx);
1693
1694     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1695         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1696             calculate_visual_weight(s, &s->plane[plane_index]);
1697         }
1698     }
1699
1700     encode_header(s);
1701     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1702     encode_blocks(s, 1);
1703     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1704
1705     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1706         Plane *p= &s->plane[plane_index];
1707         int w= p->width;
1708         int h= p->height;
1709         int x, y;
1710 //        int bits= put_bits_count(&s->c.pb);
1711
1712         if (!s->memc_only) {
1713             //FIXME optimize
1714             if(pict->data[plane_index]) //FIXME gray hack
1715                 for(y=0; y<h; y++){
1716                     for(x=0; x<w; x++){
1717                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1718                     }
1719                 }
1720             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1721
1722             if(   plane_index==0
1723                && pic->pict_type == AV_PICTURE_TYPE_P
1724                && !(avctx->flags&CODEC_FLAG_PASS2)
1725                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1726                 ff_init_range_encoder(c, pkt->data, pkt->size);
1727                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1728                 pic->pict_type= AV_PICTURE_TYPE_I;
1729                 s->keyframe=1;
1730                 s->current_picture->key_frame=1;
1731                 goto redo_frame;
1732             }
1733
1734             if(s->qlog == LOSSLESS_QLOG){
1735                 for(y=0; y<h; y++){
1736                     for(x=0; x<w; x++){
1737                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1738                     }
1739                 }
1740             }else{
1741                 for(y=0; y<h; y++){
1742                     for(x=0; x<w; x++){
1743                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1744                     }
1745                 }
1746             }
1747
1748             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1749
1750             if(s->pass1_rc && plane_index==0){
1751                 int delta_qlog = ratecontrol_1pass(s, pic);
1752                 if (delta_qlog <= INT_MIN)
1753                     return -1;
1754                 if(delta_qlog){
1755                     //reordering qlog in the bitstream would eliminate this reset
1756                     ff_init_range_encoder(c, pkt->data, pkt->size);
1757                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1758                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1759                     encode_header(s);
1760                     encode_blocks(s, 0);
1761                 }
1762             }
1763
1764             for(level=0; level<s->spatial_decomposition_count; level++){
1765                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1766                     SubBand *b= &p->band[level][orientation];
1767
1768                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1769                     if(orientation==0)
1770                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1771                     if (!s->no_bitstream)
1772                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1773                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1774                     if(orientation==0)
1775                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1776                 }
1777             }
1778
1779             for(level=0; level<s->spatial_decomposition_count; level++){
1780                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1781                     SubBand *b= &p->band[level][orientation];
1782
1783                     dequantize(s, b, b->ibuf, b->stride);
1784                 }
1785             }
1786
1787             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1788             if(s->qlog == LOSSLESS_QLOG){
1789                 for(y=0; y<h; y++){
1790                     for(x=0; x<w; x++){
1791                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1792                     }
1793                 }
1794             }
1795             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1796         }else{
1797             //ME/MC only
1798             if(pic->pict_type == AV_PICTURE_TYPE_I){
1799                 for(y=0; y<h; y++){
1800                     for(x=0; x<w; x++){
1801                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1802                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1803                     }
1804                 }
1805             }else{
1806                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1807                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1808             }
1809         }
1810         if(s->avctx->flags&CODEC_FLAG_PSNR){
1811             int64_t error= 0;
1812
1813             if(pict->data[plane_index]) //FIXME gray hack
1814                 for(y=0; y<h; y++){
1815                     for(x=0; x<w; x++){
1816                         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];
1817                         error += d*d;
1818                     }
1819                 }
1820             s->avctx->error[plane_index] += error;
1821             s->current_picture->error[plane_index] = error;
1822         }
1823
1824     }
1825
1826     update_last_header_values(s);
1827
1828     ff_snow_release_buffer(avctx);
1829
1830     s->current_picture->coded_picture_number = avctx->frame_number;
1831     s->current_picture->pict_type = pict->pict_type;
1832     s->current_picture->quality = pict->quality;
1833     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1834     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1835     s->m.current_picture.f->display_picture_number =
1836     s->m.current_picture.f->coded_picture_number   = avctx->frame_number;
1837     s->m.current_picture.f->quality                = pic->quality;
1838     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1839     if(s->pass1_rc)
1840         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1841             return -1;
1842     if(avctx->flags&CODEC_FLAG_PASS1)
1843         ff_write_pass1_stats(&s->m);
1844     s->m.last_pict_type = s->m.pict_type;
1845     avctx->frame_bits = s->m.frame_bits;
1846     avctx->mv_bits = s->m.mv_bits;
1847     avctx->misc_bits = s->m.misc_bits;
1848     avctx->p_tex_bits = s->m.p_tex_bits;
1849
1850     emms_c();
1851
1852     pkt->size = ff_rac_terminate(c);
1853     if (avctx->coded_frame->key_frame)
1854         pkt->flags |= AV_PKT_FLAG_KEY;
1855     *got_packet = 1;
1856
1857     return 0;
1858 }
1859
1860 static av_cold int encode_end(AVCodecContext *avctx)
1861 {
1862     SnowContext *s = avctx->priv_data;
1863
1864     ff_snow_common_end(s);
1865     ff_rate_control_uninit(&s->m);
1866     av_frame_free(&s->input_picture);
1867     av_free(avctx->stats_out);
1868
1869     return 0;
1870 }
1871
1872 #define OFFSET(x) offsetof(SnowContext, x)
1873 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1874 static const AVOption options[] = {
1875     { "memc_only",      "Only do ME/MC (I frames -> ref, P frame -> ME+MC).",   OFFSET(memc_only), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1876     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1877     { NULL },
1878 };
1879
1880 static const AVClass snowenc_class = {
1881     .class_name = "snow encoder",
1882     .item_name  = av_default_item_name,
1883     .option     = options,
1884     .version    = LIBAVUTIL_VERSION_INT,
1885 };
1886
1887 AVCodec ff_snow_encoder = {
1888     .name           = "snow",
1889     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1890     .type           = AVMEDIA_TYPE_VIDEO,
1891     .id             = AV_CODEC_ID_SNOW,
1892     .priv_data_size = sizeof(SnowContext),
1893     .init           = encode_init,
1894     .encode2        = encode_frame,
1895     .close          = encode_end,
1896     .pix_fmts       = (const enum AVPixelFormat[]){
1897         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1898         AV_PIX_FMT_GRAY8,
1899         AV_PIX_FMT_NONE
1900     },
1901     .priv_class     = &snowenc_class,
1902 };
1903
1904
1905 #ifdef TEST
1906 #undef malloc
1907 #undef free
1908 #undef printf
1909
1910 #include "libavutil/lfg.h"
1911 #include "libavutil/mathematics.h"
1912
1913 int main(void){
1914 #define width  256
1915 #define height 256
1916     int buffer[2][width*height];
1917     SnowContext s;
1918     int i;
1919     AVLFG prng;
1920     s.spatial_decomposition_count=6;
1921     s.spatial_decomposition_type=1;
1922
1923     s.temp_dwt_buffer  = av_mallocz(width * sizeof(DWTELEM));
1924     s.temp_idwt_buffer = av_mallocz(width * sizeof(IDWTELEM));
1925
1926     av_lfg_init(&prng, 1);
1927
1928     printf("testing 5/3 DWT\n");
1929     for(i=0; i<width*height; i++)
1930         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1931
1932     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1933     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1934
1935     for(i=0; i<width*height; i++)
1936         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1937
1938     printf("testing 9/7 DWT\n");
1939     s.spatial_decomposition_type=0;
1940     for(i=0; i<width*height; i++)
1941         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1942
1943     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1944     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1945
1946     for(i=0; i<width*height; i++)
1947         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1948
1949     {
1950     int level, orientation, x, y;
1951     int64_t errors[8][4];
1952     int64_t g=0;
1953
1954         memset(errors, 0, sizeof(errors));
1955         s.spatial_decomposition_count=3;
1956         s.spatial_decomposition_type=0;
1957         for(level=0; level<s.spatial_decomposition_count; level++){
1958             for(orientation=level ? 1 : 0; orientation<4; orientation++){
1959                 int w= width  >> (s.spatial_decomposition_count-level);
1960                 int h= height >> (s.spatial_decomposition_count-level);
1961                 int stride= width  << (s.spatial_decomposition_count-level);
1962                 DWTELEM *buf= buffer[0];
1963                 int64_t error=0;
1964
1965                 if(orientation&1) buf+=w;
1966                 if(orientation>1) buf+=stride>>1;
1967
1968                 memset(buffer[0], 0, sizeof(int)*width*height);
1969                 buf[w/2 + h/2*stride]= 256*256;
1970                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1971                 for(y=0; y<height; y++){
1972                     for(x=0; x<width; x++){
1973                         int64_t d= buffer[0][x + y*width];
1974                         error += d*d;
1975                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
1976                     }
1977                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
1978                 }
1979                 error= (int)(sqrt(error)+0.5);
1980                 errors[level][orientation]= error;
1981                 if(g) g=av_gcd(g, error);
1982                 else g= error;
1983             }
1984         }
1985         printf("static int const visual_weight[][4]={\n");
1986         for(level=0; level<s.spatial_decomposition_count; level++){
1987             printf("  {");
1988             for(orientation=0; orientation<4; orientation++){
1989                 printf("%8"PRId64",", errors[level][orientation]/g);
1990             }
1991             printf("},\n");
1992         }
1993         printf("};\n");
1994         {
1995             int level=2;
1996             int w= width  >> (s.spatial_decomposition_count-level);
1997             //int h= height >> (s.spatial_decomposition_count-level);
1998             int stride= width  << (s.spatial_decomposition_count-level);
1999             DWTELEM *buf= buffer[0];
2000             int64_t error=0;
2001
2002             buf+=w;
2003             buf+=stride>>1;
2004
2005             memset(buffer[0], 0, sizeof(int)*width*height);
2006             for(y=0; y<height; y++){
2007                 for(x=0; x<width; x++){
2008                     int tab[4]={0,2,3,1};
2009                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2010                 }
2011             }
2012             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2013             for(y=0; y<height; y++){
2014                 for(x=0; x<width; x++){
2015                     int64_t d= buffer[0][x + y*width];
2016                     error += d*d;
2017                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2018                 }
2019                 if(FFABS(height/2-y)<9) printf("\n");
2020             }
2021         }
2022
2023     }
2024     return 0;
2025 }
2026 #endif /* TEST */