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

   1 /*
   2  * Copyright (C) 2004 Michael Niedermayer <michaelni@gmx.at>
   3  *
   4  * This file is part of FFmpeg.
   5  *
   6  * FFmpeg is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * FFmpeg is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with FFmpeg; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 #include "libavutil/intmath.h"
  22 #include "libavutil/log.h"
  23 #include "libavutil/opt.h"
  24 #include "avcodec.h"
  25 #include "internal.h"
  26 #include "dsputil.h"
  27 #include "internal.h"
  28 #include "dwt.h"
  29 #include "snow.h"
  30
  31 #include "rangecoder.h"
  32 #include "mathops.h"
  33
  34 #include "mpegvideo.h"
  35 #include "h263.h"
  36
  37 #undef NDEBUG
  38 #include <assert.h>
  39
  40 #define QUANTIZE2 0
  41
  42 #if QUANTIZE2==1
  43 #define Q2_STEP 8
  44
  45 static void find_sse(SnowContext *s, Plane *p, int *score, int score_stride, IDWTELEM *r0, IDWTELEM *r1, int level, int orientation){
  46     SubBand *b= &p->band[level][orientation];
  47     int x, y;
  48     int xo=0;
  49     int yo=0;
  50     int step= 1 << (s->spatial_decomposition_count - level);
  51
  52     if(orientation&1)
  53         xo= step>>1;
  54     if(orientation&2)
  55         yo= step>>1;
  56
  57     //FIXME bias for nonzero ?
  58     //FIXME optimize
  59     memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
  60     for(y=0; y<p->height; y++){
  61         for(x=0; x<p->width; x++){
  62             int sx= (x-xo + step/2) / step / Q2_STEP;
  63             int sy= (y-yo + step/2) / step / Q2_STEP;
  64             int v= r0[x + y*p->width] - r1[x + y*p->width];
  65             assert(sx>=0 && sy>=0 && sx < score_stride);
  66             v= ((v+8)>>4)<<4;
  67             score[sx + sy*score_stride] += v*v;
  68             assert(score[sx + sy*score_stride] >= 0);
  69         }
  70     }
  71 }
  72
  73 static void dequantize_all(SnowContext *s, Plane *p, IDWTELEM *buffer, int width, int height){
  74     int level, orientation;
  75
  76     for(level=0; level<s->spatial_decomposition_count; level++){
  77         for(orientation=level ? 1 : 0; orientation<4; orientation++){
  78             SubBand *b= &p->band[level][orientation];
  79             IDWTELEM *dst= buffer + (b->ibuf - s->spatial_idwt_buffer);
  80
  81             dequantize(s, b, dst, b->stride);
  82         }
  83     }
  84 }
  85
  86 static void dwt_quantize(SnowContext *s, Plane *p, DWTELEM *buffer, int width, int height, int stride, int type){
  87     int level, orientation, ys, xs, x, y, pass;
  88     IDWTELEM best_dequant[height * stride];
  89     IDWTELEM idwt2_buffer[height * stride];
  90     const int score_stride= (width + 10)/Q2_STEP;
  91     int best_score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
  92     int score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
  93     int threshold= (s->m.lambda * s->m.lambda) >> 6;
  94
  95     //FIXME pass the copy cleanly ?
  96
  97 //    memcpy(dwt_buffer, buffer, height * stride * sizeof(DWTELEM));
  98     ff_spatial_dwt(buffer, s->temp_dwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
  99
 100     for(level=0; level<s->spatial_decomposition_count; level++){
 101         for(orientation=level ? 1 : 0; orientation<4; orientation++){
 102             SubBand *b= &p->band[level][orientation];
 103             IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
 104              DWTELEM *src=       buffer + (b-> buf - s->spatial_dwt_buffer);
 105             assert(src == b->buf); // code does not depend on this but it is true currently
 106
 107             quantize(s, b, dst, src, b->stride, s->qbias);
 108         }
 109     }
 110     for(pass=0; pass<1; pass++){
 111         if(s->qbias == 0) //keyframe
 112             continue;
 113         for(level=0; level<s->spatial_decomposition_count; level++){
 114             for(orientation=level ? 1 : 0; orientation<4; orientation++){
 115                 SubBand *b= &p->band[level][orientation];
 116                 IDWTELEM *dst= idwt2_buffer + (b->ibuf - s->spatial_idwt_buffer);
 117                 IDWTELEM *best_dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
 118
 119                 for(ys= 0; ys<Q2_STEP; ys++){
 120                     for(xs= 0; xs<Q2_STEP; xs++){
 121                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
 122                         dequantize_all(s, p, idwt2_buffer, width, height);
 123                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
 124                         find_sse(s, p, best_score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
 125                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
 126                         for(y=ys; y<b->height; y+= Q2_STEP){
 127                             for(x=xs; x<b->width; x+= Q2_STEP){
 128                                 if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
 129                                 if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
 130                                 //FIXME try more than just --
 131                             }
 132                         }
 133                         dequantize_all(s, p, idwt2_buffer, width, height);
 134                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
 135                         find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
 136                         for(y=ys; y<b->height; y+= Q2_STEP){
 137                             for(x=xs; x<b->width; x+= Q2_STEP){
 138                                 int score_idx= x/Q2_STEP + (y/Q2_STEP)*score_stride;
 139                                 if(score[score_idx] <= best_score[score_idx] + threshold){
 140                                     best_score[score_idx]= score[score_idx];
 141                                     if(best_dst[x + y*b->stride]<0) best_dst[x + y*b->stride]++;
 142                                     if(best_dst[x + y*b->stride]>0) best_dst[x + y*b->stride]--;
 143                                     //FIXME copy instead
 144                                 }
 145                             }
 146                         }
 147                     }
 148                 }
 149             }
 150         }
 151     }
 152     memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly instead of copy at the end
 153 }
 154
 155 #endif /* QUANTIZE2==1 */
 156
 157 static av_cold int encode_init(AVCodecContext *avctx)
 158 {
 159     SnowContext *s = avctx->priv_data;
 160     int plane_index, ret;
 161
 162     if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){
 163         av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"
 164                "Use vstrict=-2 / -strict -2 to use it anyway.\n");
 165         return -1;
 166     }
 167
 168     if(avctx->prediction_method == DWT_97
 169        && (avctx->flags & CODEC_FLAG_QSCALE)
 170        && avctx->global_quality == 0){
 171         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
 172         return -1;
 173     }
 174
 175     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
 176
 177     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
 178     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
 179
 180     for(plane_index=0; plane_index<3; plane_index++){
 181         s->plane[plane_index].diag_mc= 1;
 182         s->plane[plane_index].htaps= 6;
 183         s->plane[plane_index].hcoeff[0]=  40;
 184         s->plane[plane_index].hcoeff[1]= -10;
 185         s->plane[plane_index].hcoeff[2]=   2;
 186         s->plane[plane_index].fast_mc= 1;
 187     }
 188
 189     if ((ret = ff_snow_common_init(avctx)) < 0) {
 190         ff_snow_common_end(avctx->priv_data);
 191         return ret;
 192     }
 193     ff_snow_alloc_blocks(s);
 194
 195     s->version=0;
 196
 197     s->m.avctx   = avctx;
 198     s->m.flags   = avctx->flags;
 199     s->m.bit_rate= avctx->bit_rate;
 200
 201     s->m.me.temp      =
 202     s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
 203     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
 204     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
 205     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
 206     ff_h263_encode_init(&s->m); //mv_penalty
 207
 208     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
 209
 210     if(avctx->flags&CODEC_FLAG_PASS1){
 211         if(!avctx->stats_out)
 212             avctx->stats_out = av_mallocz(256);
 213     }
 214     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
 215         if(ff_rate_control_init(&s->m) < 0)
 216             return -1;
 217     }
 218     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
 219
 220     avctx->coded_frame= &s->current_picture;
 221     switch(avctx->pix_fmt){
 222     case AV_PIX_FMT_YUV444P:
 223 //    case AV_PIX_FMT_YUV422P:
 224     case AV_PIX_FMT_YUV420P:
 225 //     case AV_PIX_FMT_GRAY8:
 226 //    case AV_PIX_FMT_YUV411P:
 227     case AV_PIX_FMT_YUV410P:
 228         s->colorspace_type= 0;
 229         break;
 230 /*    case AV_PIX_FMT_RGB32:
 231         s->colorspace= 1;
 232         break;*/
 233     default:
 234         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
 235         return -1;
 236     }
 237     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
 238
 239     ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
 240     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
 241
 242     ff_get_buffer(s->avctx, &s->input_picture);
 243
 244     if(s->avctx->me_method == ME_ITER){
 245         int i;
 246         int size= s->b_width * s->b_height << 2*s->block_max_depth;
 247         for(i=0; i<s->max_ref_frames; i++){
 248             s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2]));
 249             s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t));
 250         }
 251     }
 252
 253     return 0;
 254 }
 255
 256 //near copy & paste from dsputil, FIXME
 257 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
 258 {
 259     int s, i, j;
 260
 261     s = 0;
 262     for (i = 0; i < h; i++) {
 263         for (j = 0; j < w; j++) {
 264             s += pix[0];
 265             pix ++;
 266         }
 267         pix += line_size - w;
 268     }
 269     return s;
 270 }
 271
 272 //near copy & paste from dsputil, FIXME
 273 static int pix_norm1(uint8_t * pix, int line_size, int w)
 274 {
 275     int s, i, j;
 276     uint32_t *sq = ff_squareTbl + 256;
 277
 278     s = 0;
 279     for (i = 0; i < w; i++) {
 280         for (j = 0; j < w; j ++) {
 281             s += sq[pix[0]];
 282             pix ++;
 283         }
 284         pix += line_size - w;
 285     }
 286     return s;
 287 }
 288
 289 //FIXME copy&paste
 290 #define P_LEFT P[1]
 291 #define P_TOP P[2]
 292 #define P_TOPRIGHT P[3]
 293 #define P_MEDIAN P[4]
 294 #define P_MV1 P[9]
 295 #define FLAG_QPEL   1 //must be 1
 296
 297 static int encode_q_branch(SnowContext *s, int level, int x, int y){
 298     uint8_t p_buffer[1024];
 299     uint8_t i_buffer[1024];
 300     uint8_t p_state[sizeof(s->block_state)];
 301     uint8_t i_state[sizeof(s->block_state)];
 302     RangeCoder pc, ic;
 303     uint8_t *pbbak= s->c.bytestream;
 304     uint8_t *pbbak_start= s->c.bytestream_start;
 305     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
 306     const int w= s->b_width  << s->block_max_depth;
 307     const int h= s->b_height << s->block_max_depth;
 308     const int rem_depth= s->block_max_depth - level;
 309     const int index= (x + y*w) << rem_depth;
 310     const int block_w= 1<<(LOG2_MB_SIZE - level);
 311     int trx= (x+1)<<rem_depth;
 312     int try= (y+1)<<rem_depth;
 313     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 314     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 315     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
 316     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
 317     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 318     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 319     int pl = left->color[0];
 320     int pcb= left->color[1];
 321     int pcr= left->color[2];
 322     int pmx, pmy;
 323     int mx=0, my=0;
 324     int l,cr,cb;
 325     const int stride= s->current_picture.linesize[0];
 326     const int uvstride= s->current_picture.linesize[1];
 327     uint8_t *current_data[3]= { s->input_picture.data[0] + (x + y*  stride)*block_w,
 328                                 s->input_picture.data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
 329                                 s->input_picture.data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
 330     int P[10][2];
 331     int16_t last_mv[3][2];
 332     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
 333     const int shift= 1+qpel;
 334     MotionEstContext *c= &s->m.me;
 335     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 336     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 337     int my_context= av_log2(2*FFABS(left->my - top->my));
 338     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 339     int ref, best_ref, ref_score, ref_mx, ref_my;
 340
 341     assert(sizeof(s->block_state) >= 256);
 342     if(s->keyframe){
 343         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 344         return 0;
 345     }
 346
 347 //    clip predictors / edge ?
 348
 349     P_LEFT[0]= left->mx;
 350     P_LEFT[1]= left->my;
 351     P_TOP [0]= top->mx;
 352     P_TOP [1]= top->my;
 353     P_TOPRIGHT[0]= tr->mx;
 354     P_TOPRIGHT[1]= tr->my;
 355
 356     last_mv[0][0]= s->block[index].mx;
 357     last_mv[0][1]= s->block[index].my;
 358     last_mv[1][0]= right->mx;
 359     last_mv[1][1]= right->my;
 360     last_mv[2][0]= bottom->mx;
 361     last_mv[2][1]= bottom->my;
 362
 363     s->m.mb_stride=2;
 364     s->m.mb_x=
 365     s->m.mb_y= 0;
 366     c->skip= 0;
 367
 368     assert(c->  stride ==   stride);
 369     assert(c->uvstride == uvstride);
 370
 371     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
 372     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
 373     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
 374     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
 375
 376     c->xmin = - x*block_w - 16+3;
 377     c->ymin = - y*block_w - 16+3;
 378     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 379     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
 380
 381     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
 382     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
 383     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
 384     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
 385     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
 386     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
 387     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
 388
 389     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
 390     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 391
 392     if (!y) {
 393         c->pred_x= P_LEFT[0];
 394         c->pred_y= P_LEFT[1];
 395     } else {
 396         c->pred_x = P_MEDIAN[0];
 397         c->pred_y = P_MEDIAN[1];
 398     }
 399
 400     score= INT_MAX;
 401     best_ref= 0;
 402     for(ref=0; ref<s->ref_frames; ref++){
 403         init_ref(c, current_data, s->last_picture[ref].data, NULL, block_w*x, block_w*y, 0);
 404
 405         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
 406                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
 407
 408         assert(ref_mx >= c->xmin);
 409         assert(ref_mx <= c->xmax);
 410         assert(ref_my >= c->ymin);
 411         assert(ref_my <= c->ymax);
 412
 413         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
 414         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
 415         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
 416         if(s->ref_mvs[ref]){
 417             s->ref_mvs[ref][index][0]= ref_mx;
 418             s->ref_mvs[ref][index][1]= ref_my;
 419             s->ref_scores[ref][index]= ref_score;
 420         }
 421         if(score > ref_score){
 422             score= ref_score;
 423             best_ref= ref;
 424             mx= ref_mx;
 425             my= ref_my;
 426         }
 427     }
 428     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
 429
 430   //  subpel search
 431     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
 432     pc= s->c;
 433     pc.bytestream_start=
 434     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
 435     memcpy(p_state, s->block_state, sizeof(s->block_state));
 436
 437     if(level!=s->block_max_depth)
 438         put_rac(&pc, &p_state[4 + s_context], 1);
 439     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
 440     if(s->ref_frames > 1)
 441         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
 442     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
 443     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
 444     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
 445     p_len= pc.bytestream - pc.bytestream_start;
 446     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
 447
 448     block_s= block_w*block_w;
 449     sum = pix_sum(current_data[0], stride, block_w, block_w);
 450     l= (sum + block_s/2)/block_s;
 451     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
 452
 453     block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
 454     sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 455     cb= (sum + block_s/2)/block_s;
 456 //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
 457     sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
 458     cr= (sum + block_s/2)/block_s;
 459 //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
 460
 461     ic= s->c;
 462     ic.bytestream_start=
 463     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
 464     memcpy(i_state, s->block_state, sizeof(s->block_state));
 465     if(level!=s->block_max_depth)
 466         put_rac(&ic, &i_state[4 + s_context], 1);
 467     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
 468     put_symbol(&ic, &i_state[32],  l-pl , 1);
 469     put_symbol(&ic, &i_state[64], cb-pcb, 1);
 470     put_symbol(&ic, &i_state[96], cr-pcr, 1);
 471     i_len= ic.bytestream - ic.bytestream_start;
 472     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
 473
 474 //    assert(score==256*256*256*64-1);
 475     assert(iscore < 255*255*256 + s->lambda2*10);
 476     assert(iscore >= 0);
 477     assert(l>=0 && l<=255);
 478     assert(pl>=0 && pl<=255);
 479
 480     if(level==0){
 481         int varc= iscore >> 8;
 482         int vard= score >> 8;
 483         if (vard <= 64 || vard < varc)
 484             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
 485         else
 486             c->scene_change_score+= s->m.qscale;
 487     }
 488
 489     if(level!=s->block_max_depth){
 490         put_rac(&s->c, &s->block_state[4 + s_context], 0);
 491         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
 492         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
 493         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
 494         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
 495         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
 496
 497         if(score2 < score && score2 < iscore)
 498             return score2;
 499     }
 500
 501     if(iscore < score){
 502         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 503         memcpy(pbbak, i_buffer, i_len);
 504         s->c= ic;
 505         s->c.bytestream_start= pbbak_start;
 506         s->c.bytestream= pbbak + i_len;
 507         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
 508         memcpy(s->block_state, i_state, sizeof(s->block_state));
 509         return iscore;
 510     }else{
 511         memcpy(pbbak, p_buffer, p_len);
 512         s->c= pc;
 513         s->c.bytestream_start= pbbak_start;
 514         s->c.bytestream= pbbak + p_len;
 515         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
 516         memcpy(s->block_state, p_state, sizeof(s->block_state));
 517         return score;
 518     }
 519 }
 520
 521 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
 522     const int w= s->b_width  << s->block_max_depth;
 523     const int rem_depth= s->block_max_depth - level;
 524     const int index= (x + y*w) << rem_depth;
 525     int trx= (x+1)<<rem_depth;
 526     BlockNode *b= &s->block[index];
 527     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 528     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
 529     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
 530     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
 531     int pl = left->color[0];
 532     int pcb= left->color[1];
 533     int pcr= left->color[2];
 534     int pmx, pmy;
 535     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
 536     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
 537     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
 538     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
 539
 540     if(s->keyframe){
 541         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
 542         return;
 543     }
 544
 545     if(level!=s->block_max_depth){
 546         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
 547             put_rac(&s->c, &s->block_state[4 + s_context], 1);
 548         }else{
 549             put_rac(&s->c, &s->block_state[4 + s_context], 0);
 550             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
 551             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
 552             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
 553             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
 554             return;
 555         }
 556     }
 557     if(b->type & BLOCK_INTRA){
 558         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
 559         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
 560         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
 561         put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
 562         put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
 563         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
 564     }else{
 565         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
 566         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
 567         if(s->ref_frames > 1)
 568             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
 569         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
 570         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
 571         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
 572     }
 573 }
 574
 575 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
 576     int i, x2, y2;
 577     Plane *p= &s->plane[plane_index];
 578     const int block_size = MB_SIZE >> s->block_max_depth;
 579     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 580     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 581     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 582     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 583     const int ref_stride= s->current_picture.linesize[plane_index];
 584     uint8_t *src= s-> input_picture.data[plane_index];
 585     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
 586     const int b_stride = s->b_width << s->block_max_depth;
 587     const int w= p->width;
 588     const int h= p->height;
 589     int index= mb_x + mb_y*b_stride;
 590     BlockNode *b= &s->block[index];
 591     BlockNode backup= *b;
 592     int ab=0;
 593     int aa=0;
 594
 595     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
 596
 597     b->type|= BLOCK_INTRA;
 598     b->color[plane_index]= 0;
 599     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
 600
 601     for(i=0; i<4; i++){
 602         int mb_x2= mb_x + (i &1) - 1;
 603         int mb_y2= mb_y + (i>>1) - 1;
 604         int x= block_w*mb_x2 + block_w/2;
 605         int y= block_h*mb_y2 + block_h/2;
 606
 607         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
 608                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
 609
 610         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
 611             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
 612                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
 613                 int obmc_v= obmc[index];
 614                 int d;
 615                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
 616                 if(x<0) obmc_v += obmc[index + block_w];
 617                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
 618                 if(x+block_w>w) obmc_v += obmc[index - block_w];
 619                 //FIXME precalculate this or simplify it somehow else
 620
 621                 d = -dst[index] + (1<<(FRAC_BITS-1));
 622                 dst[index] = d;
 623                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
 624                 aa += obmc_v * obmc_v; //FIXME precalculate this
 625             }
 626         }
 627     }
 628     *b= backup;
 629
 630     return av_clip( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa), 0, 255); //FIXME we should not need clipping
 631 }
 632
 633 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
 634     const int b_stride = s->b_width << s->block_max_depth;
 635     const int b_height = s->b_height<< s->block_max_depth;
 636     int index= x + y*b_stride;
 637     const BlockNode *b     = &s->block[index];
 638     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 639     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
 640     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
 641     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
 642     int dmx, dmy;
 643 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 644 //  int my_context= av_log2(2*FFABS(left->my - top->my));
 645
 646     if(x<0 || x>=b_stride || y>=b_height)
 647         return 0;
 648 /*
 649 1            0      0
 650 01X          1-2    1
 651 001XX        3-6    2-3
 652 0001XXX      7-14   4-7
 653 00001XXXX   15-30   8-15
 654 */
 655 //FIXME try accurate rate
 656 //FIXME intra and inter predictors if surrounding blocks are not the same type
 657     if(b->type & BLOCK_INTRA){
 658         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
 659                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
 660                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
 661     }else{
 662         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
 663         dmx-= b->mx;
 664         dmy-= b->my;
 665         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
 666                     + av_log2(2*FFABS(dmy))
 667                     + av_log2(2*b->ref));
 668     }
 669 }
 670
 671 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
 672     Plane *p= &s->plane[plane_index];
 673     const int block_size = MB_SIZE >> s->block_max_depth;
 674     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 675     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 676     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 677     const int ref_stride= s->current_picture.linesize[plane_index];
 678     uint8_t *dst= s->current_picture.data[plane_index];
 679     uint8_t *src= s->  input_picture.data[plane_index];
 680     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
 681     uint8_t *cur = s->scratchbuf;
 682     uint8_t *tmp = s->emu_edge_buffer;
 683     const int b_stride = s->b_width << s->block_max_depth;
 684     const int b_height = s->b_height<< s->block_max_depth;
 685     const int w= p->width;
 686     const int h= p->height;
 687     int distortion;
 688     int rate= 0;
 689     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 690     int sx= block_w*mb_x - block_w/2;
 691     int sy= block_h*mb_y - block_h/2;
 692     int x0= FFMAX(0,-sx);
 693     int y0= FFMAX(0,-sy);
 694     int x1= FFMIN(block_w*2, w-sx);
 695     int y1= FFMIN(block_h*2, h-sy);
 696     int i,x,y;
 697
 698     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
 699
 700     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);
 701
 702     for(y=y0; y<y1; y++){
 703         const uint8_t *obmc1= obmc_edged[y];
 704         const IDWTELEM *pred1 = pred + y*obmc_stride;
 705         uint8_t *cur1 = cur + y*ref_stride;
 706         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
 707         for(x=x0; x<x1; x++){
 708 #if FRAC_BITS >= LOG2_OBMC_MAX
 709             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
 710 #else
 711             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
 712 #endif
 713             v = (v + pred1[x]) >> FRAC_BITS;
 714             if(v&(~255)) v= ~(v>>31);
 715             dst1[x] = v;
 716         }
 717     }
 718
 719     /* copy the regions where obmc[] = (uint8_t)256 */
 720     if(LOG2_OBMC_MAX == 8
 721         && (mb_x == 0 || mb_x == b_stride-1)
 722         && (mb_y == 0 || mb_y == b_height-1)){
 723         if(mb_x == 0)
 724             x1 = block_w;
 725         else
 726             x0 = block_w;
 727         if(mb_y == 0)
 728             y1 = block_h;
 729         else
 730             y0 = block_h;
 731         for(y=y0; y<y1; y++)
 732             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
 733     }
 734
 735     if(block_w==16){
 736         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
 737         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
 738         /* FIXME cmps overlap but do not cover the wavelet's whole support.
 739          * So improving the score of one block is not strictly guaranteed
 740          * to improve the score of the whole frame, thus iterative motion
 741          * estimation does not always converge. */
 742         if(s->avctx->me_cmp == FF_CMP_W97)
 743             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 744         else if(s->avctx->me_cmp == FF_CMP_W53)
 745             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 746         else{
 747             distortion = 0;
 748             for(i=0; i<4; i++){
 749                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
 750                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
 751             }
 752         }
 753     }else{
 754         assert(block_w==8);
 755         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
 756     }
 757
 758     if(plane_index==0){
 759         for(i=0; i<4; i++){
 760 /* ..RRr
 761  * .RXx.
 762  * rxx..
 763  */
 764             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
 765         }
 766         if(mb_x == b_stride-2)
 767             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
 768     }
 769     return distortion + rate*penalty_factor;
 770 }
 771
 772 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
 773     int i, y2;
 774     Plane *p= &s->plane[plane_index];
 775     const int block_size = MB_SIZE >> s->block_max_depth;
 776     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
 777     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
 778     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
 779     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
 780     const int ref_stride= s->current_picture.linesize[plane_index];
 781     uint8_t *dst= s->current_picture.data[plane_index];
 782     uint8_t *src= s-> input_picture.data[plane_index];
 783     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
 784     // const has only been removed from zero_dst to suppress a warning
 785     static IDWTELEM zero_dst[4096]; //FIXME
 786     const int b_stride = s->b_width << s->block_max_depth;
 787     const int w= p->width;
 788     const int h= p->height;
 789     int distortion= 0;
 790     int rate= 0;
 791     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 792
 793     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
 794
 795     for(i=0; i<9; i++){
 796         int mb_x2= mb_x + (i%3) - 1;
 797         int mb_y2= mb_y + (i/3) - 1;
 798         int x= block_w*mb_x2 + block_w/2;
 799         int y= block_h*mb_y2 + block_h/2;
 800
 801         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
 802                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
 803
 804         //FIXME find a cleaner/simpler way to skip the outside stuff
 805         for(y2= y; y2<0; y2++)
 806             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 807         for(y2= h; y2<y+block_h; y2++)
 808             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 809         if(x<0){
 810             for(y2= y; y2<y+block_h; y2++)
 811                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
 812         }
 813         if(x+block_w > w){
 814             for(y2= y; y2<y+block_h; y2++)
 815                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
 816         }
 817
 818         assert(block_w== 8 || block_w==16);
 819         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
 820     }
 821
 822     if(plane_index==0){
 823         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
 824         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
 825
 826 /* ..RRRr
 827  * .RXXx.
 828  * .RXXx.
 829  * rxxx.
 830  */
 831         if(merged)
 832             rate = get_block_bits(s, mb_x, mb_y, 2);
 833         for(i=merged?4:0; i<9; i++){
 834             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
 835             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
 836         }
 837     }
 838     return distortion + rate*penalty_factor;
 839 }
 840
 841 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 842     const int w= b->width;
 843     const int h= b->height;
 844     int x, y;
 845
 846     if(1){
 847         int run=0;
 848         int *runs = s->run_buffer;
 849         int run_index=0;
 850         int max_index;
 851
 852         for(y=0; y<h; y++){
 853             for(x=0; x<w; x++){
 854                 int v, p=0;
 855                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 856                 v= src[x + y*stride];
 857
 858                 if(y){
 859                     t= src[x + (y-1)*stride];
 860                     if(x){
 861                         lt= src[x - 1 + (y-1)*stride];
 862                     }
 863                     if(x + 1 < w){
 864                         rt= src[x + 1 + (y-1)*stride];
 865                     }
 866                 }
 867                 if(x){
 868                     l= src[x - 1 + y*stride];
 869                     /*if(x > 1){
 870                         if(orientation==1) ll= src[y + (x-2)*stride];
 871                         else               ll= src[x - 2 + y*stride];
 872                     }*/
 873                 }
 874                 if(parent){
 875                     int px= x>>1;
 876                     int py= y>>1;
 877                     if(px<b->parent->width && py<b->parent->height)
 878                         p= parent[px + py*2*stride];
 879                 }
 880                 if(!(/*ll|*/l|lt|t|rt|p)){
 881                     if(v){
 882                         runs[run_index++]= run;
 883                         run=0;
 884                     }else{
 885                         run++;
 886                     }
 887                 }
 888             }
 889         }
 890         max_index= run_index;
 891         runs[run_index++]= run;
 892         run_index=0;
 893         run= runs[run_index++];
 894
 895         put_symbol2(&s->c, b->state[30], max_index, 0);
 896         if(run_index <= max_index)
 897             put_symbol2(&s->c, b->state[1], run, 3);
 898
 899         for(y=0; y<h; y++){
 900             if(s->c.bytestream_end - s->c.bytestream < w*40){
 901                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 902                 return -1;
 903             }
 904             for(x=0; x<w; x++){
 905                 int v, p=0;
 906                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 907                 v= src[x + y*stride];
 908
 909                 if(y){
 910                     t= src[x + (y-1)*stride];
 911                     if(x){
 912                         lt= src[x - 1 + (y-1)*stride];
 913                     }
 914                     if(x + 1 < w){
 915                         rt= src[x + 1 + (y-1)*stride];
 916                     }
 917                 }
 918                 if(x){
 919                     l= src[x - 1 + y*stride];
 920                     /*if(x > 1){
 921                         if(orientation==1) ll= src[y + (x-2)*stride];
 922                         else               ll= src[x - 2 + y*stride];
 923                     }*/
 924                 }
 925                 if(parent){
 926                     int px= x>>1;
 927                     int py= y>>1;
 928                     if(px<b->parent->width && py<b->parent->height)
 929                         p= parent[px + py*2*stride];
 930                 }
 931                 if(/*ll|*/l|lt|t|rt|p){
 932                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 933
 934                     put_rac(&s->c, &b->state[0][context], !!v);
 935                 }else{
 936                     if(!run){
 937                         run= runs[run_index++];
 938
 939                         if(run_index <= max_index)
 940                             put_symbol2(&s->c, b->state[1], run, 3);
 941                         assert(v);
 942                     }else{
 943                         run--;
 944                         assert(!v);
 945                     }
 946                 }
 947                 if(v){
 948                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 949                     int l2= 2*FFABS(l) + (l<0);
 950                     int t2= 2*FFABS(t) + (t<0);
 951
 952                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
 953                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
 954                 }
 955             }
 956         }
 957     }
 958     return 0;
 959 }
 960
 961 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
 962 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
 963 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
 964     return encode_subband_c0run(s, b, src, parent, stride, orientation);
 965 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
 966 }
 967
 968 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){
 969     const int b_stride= s->b_width << s->block_max_depth;
 970     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 971     BlockNode backup= *block;
 972     unsigned value;
 973     int rd, index;
 974
 975     assert(mb_x>=0 && mb_y>=0);
 976     assert(mb_x<b_stride);
 977
 978     if(intra){
 979         block->color[0] = p[0];
 980         block->color[1] = p[1];
 981         block->color[2] = p[2];
 982         block->type |= BLOCK_INTRA;
 983     }else{
 984         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
 985         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
 986         if(s->me_cache[index] == value)
 987             return 0;
 988         s->me_cache[index]= value;
 989
 990         block->mx= p[0];
 991         block->my= p[1];
 992         block->type &= ~BLOCK_INTRA;
 993     }
 994
 995     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
 996
 997 //FIXME chroma
 998     if(rd < *best_rd){
 999         *best_rd= rd;
1000         return 1;
1001     }else{
1002         *block= backup;
1003         return 0;
1004     }
1005 }
1006
1007 /* special case for int[2] args we discard afterwards,
1008  * fixes compilation problem with gcc 2.95 */
1009 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){
1010     int p[2] = {p0, p1};
1011     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
1012 }
1013
1014 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){
1015     const int b_stride= s->b_width << s->block_max_depth;
1016     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
1017     BlockNode backup[4];
1018     unsigned value;
1019     int rd, index;
1020
1021     /* We don't initialize backup[] during variable declaration, because
1022      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
1023      * 'int16_t'". */
1024     backup[0] = block[0];
1025     backup[1] = block[1];
1026     backup[2] = block[b_stride];
1027     backup[3] = block[b_stride + 1];
1028
1029     assert(mb_x>=0 && mb_y>=0);
1030     assert(mb_x<b_stride);
1031     assert(((mb_x|mb_y)&1) == 0);
1032
1033     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
1034     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
1035     if(s->me_cache[index] == value)
1036         return 0;
1037     s->me_cache[index]= value;
1038
1039     block->mx= p0;
1040     block->my= p1;
1041     block->ref= ref;
1042     block->type &= ~BLOCK_INTRA;
1043     block[1]= block[b_stride]= block[b_stride+1]= *block;
1044
1045     rd= get_4block_rd(s, mb_x, mb_y, 0);
1046
1047 //FIXME chroma
1048     if(rd < *best_rd){
1049         *best_rd= rd;
1050         return 1;
1051     }else{
1052         block[0]= backup[0];
1053         block[1]= backup[1];
1054         block[b_stride]= backup[2];
1055         block[b_stride+1]= backup[3];
1056         return 0;
1057     }
1058 }
1059
1060 static void iterative_me(SnowContext *s){
1061     int pass, mb_x, mb_y;
1062     const int b_width = s->b_width  << s->block_max_depth;
1063     const int b_height= s->b_height << s->block_max_depth;
1064     const int b_stride= b_width;
1065     int color[3];
1066
1067     {
1068         RangeCoder r = s->c;
1069         uint8_t state[sizeof(s->block_state)];
1070         memcpy(state, s->block_state, sizeof(s->block_state));
1071         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1072             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1073                 encode_q_branch(s, 0, mb_x, mb_y);
1074         s->c = r;
1075         memcpy(s->block_state, state, sizeof(s->block_state));
1076     }
1077
1078     for(pass=0; pass<25; pass++){
1079         int change= 0;
1080
1081         for(mb_y= 0; mb_y<b_height; mb_y++){
1082             for(mb_x= 0; mb_x<b_width; mb_x++){
1083                 int dia_change, i, j, ref;
1084                 int best_rd= INT_MAX, ref_rd;
1085                 BlockNode backup, ref_b;
1086                 const int index= mb_x + mb_y * b_stride;
1087                 BlockNode *block= &s->block[index];
1088                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1089                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1090                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1091                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1092                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1093                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1094                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1095                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1096                 const int b_w= (MB_SIZE >> s->block_max_depth);
1097                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1098
1099                 if(pass && (block->type & BLOCK_OPT))
1100                     continue;
1101                 block->type |= BLOCK_OPT;
1102
1103                 backup= *block;
1104
1105                 if(!s->me_cache_generation)
1106                     memset(s->me_cache, 0, sizeof(s->me_cache));
1107                 s->me_cache_generation += 1<<22;
1108
1109                 //FIXME precalculate
1110                 {
1111                     int x, y;
1112                     for (y = 0; y < b_w * 2; y++)
1113                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1114                     if(mb_x==0)
1115                         for(y=0; y<b_w*2; y++)
1116                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1117                     if(mb_x==b_stride-1)
1118                         for(y=0; y<b_w*2; y++)
1119                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1120                     if(mb_y==0){
1121                         for(x=0; x<b_w*2; x++)
1122                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1123                         for(y=1; y<b_w; y++)
1124                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1125                     }
1126                     if(mb_y==b_height-1){
1127                         for(x=0; x<b_w*2; x++)
1128                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1129                         for(y=b_w; y<b_w*2-1; y++)
1130                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1131                     }
1132                 }
1133
1134                 //skip stuff outside the picture
1135                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1136                     uint8_t *src= s->  input_picture.data[0];
1137                     uint8_t *dst= s->current_picture.data[0];
1138                     const int stride= s->current_picture.linesize[0];
1139                     const int block_w= MB_SIZE >> s->block_max_depth;
1140                     const int block_h= MB_SIZE >> s->block_max_depth;
1141                     const int sx= block_w*mb_x - block_w/2;
1142                     const int sy= block_h*mb_y - block_h/2;
1143                     const int w= s->plane[0].width;
1144                     const int h= s->plane[0].height;
1145                     int y;
1146
1147                     for(y=sy; y<0; y++)
1148                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1149                     for(y=h; y<sy+block_h*2; y++)
1150                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1151                     if(sx<0){
1152                         for(y=sy; y<sy+block_h*2; y++)
1153                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1154                     }
1155                     if(sx+block_w*2 > w){
1156                         for(y=sy; y<sy+block_h*2; y++)
1157                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1158                     }
1159                 }
1160
1161                 // intra(black) = neighbors' contribution to the current block
1162                 for(i=0; i<3; i++)
1163                     color[i]= get_dc(s, mb_x, mb_y, i);
1164
1165                 // get previous score (cannot be cached due to OBMC)
1166                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1167                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1168                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1169                 }else
1170                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1171
1172                 ref_b= *block;
1173                 ref_rd= best_rd;
1174                 for(ref=0; ref < s->ref_frames; ref++){
1175                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1176                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1177                         continue;
1178                     block->ref= ref;
1179                     best_rd= INT_MAX;
1180
1181                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1182                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1183                     if(tb)
1184                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1185                     if(lb)
1186                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1187                     if(rb)
1188                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1189                     if(bb)
1190                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1191
1192                     /* fullpel ME */
1193                     //FIXME avoid subpel interpolation / round to nearest integer
1194                     do{
1195                         dia_change=0;
1196                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1197                             for(j=0; j<i; j++){
1198                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1199                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1200                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1201                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1202                             }
1203                         }
1204                     }while(dia_change);
1205                     /* subpel ME */
1206                     do{
1207                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1208                         dia_change=0;
1209                         for(i=0; i<8; i++)
1210                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1211                     }while(dia_change);
1212                     //FIXME or try the standard 2 pass qpel or similar
1213
1214                     mvr[0][0]= block->mx;
1215                     mvr[0][1]= block->my;
1216                     if(ref_rd > best_rd){
1217                         ref_rd= best_rd;
1218                         ref_b= *block;
1219                     }
1220                 }
1221                 best_rd= ref_rd;
1222                 *block= ref_b;
1223                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1224                 //FIXME RD style color selection
1225                 if(!same_block(block, &backup)){
1226                     if(tb ) tb ->type &= ~BLOCK_OPT;
1227                     if(lb ) lb ->type &= ~BLOCK_OPT;
1228                     if(rb ) rb ->type &= ~BLOCK_OPT;
1229                     if(bb ) bb ->type &= ~BLOCK_OPT;
1230                     if(tlb) tlb->type &= ~BLOCK_OPT;
1231                     if(trb) trb->type &= ~BLOCK_OPT;
1232                     if(blb) blb->type &= ~BLOCK_OPT;
1233                     if(brb) brb->type &= ~BLOCK_OPT;
1234                     change ++;
1235                 }
1236             }
1237         }
1238         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1239         if(!change)
1240             break;
1241     }
1242
1243     if(s->block_max_depth == 1){
1244         int change= 0;
1245         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1246             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1247                 int i;
1248                 int best_rd, init_rd;
1249                 const int index= mb_x + mb_y * b_stride;
1250                 BlockNode *b[4];
1251
1252                 b[0]= &s->block[index];
1253                 b[1]= b[0]+1;
1254                 b[2]= b[0]+b_stride;
1255                 b[3]= b[2]+1;
1256                 if(same_block(b[0], b[1]) &&
1257                    same_block(b[0], b[2]) &&
1258                    same_block(b[0], b[3]))
1259                     continue;
1260
1261                 if(!s->me_cache_generation)
1262                     memset(s->me_cache, 0, sizeof(s->me_cache));
1263                 s->me_cache_generation += 1<<22;
1264
1265                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1266
1267                 //FIXME more multiref search?
1268                 check_4block_inter(s, mb_x, mb_y,
1269                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1270                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1271
1272                 for(i=0; i<4; i++)
1273                     if(!(b[i]->type&BLOCK_INTRA))
1274                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1275
1276                 if(init_rd != best_rd)
1277                     change++;
1278             }
1279         }
1280         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1281     }
1282 }
1283
1284 static void encode_blocks(SnowContext *s, int search){
1285     int x, y;
1286     int w= s->b_width;
1287     int h= s->b_height;
1288
1289     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1290         iterative_me(s);
1291
1292     for(y=0; y<h; y++){
1293         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1294             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1295             return;
1296         }
1297         for(x=0; x<w; x++){
1298             if(s->avctx->me_method == ME_ITER || !search)
1299                 encode_q_branch2(s, 0, x, y);
1300             else
1301                 encode_q_branch (s, 0, x, y);
1302         }
1303     }
1304 }
1305
1306 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1307     const int w= b->width;
1308     const int h= b->height;
1309     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1310     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1311     int x,y, thres1, thres2;
1312
1313     if(s->qlog == LOSSLESS_QLOG){
1314         for(y=0; y<h; y++)
1315             for(x=0; x<w; x++)
1316                 dst[x + y*stride]= src[x + y*stride];
1317         return;
1318     }
1319
1320     bias= bias ? 0 : (3*qmul)>>3;
1321     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1322     thres2= 2*thres1;
1323
1324     if(!bias){
1325         for(y=0; y<h; y++){
1326             for(x=0; x<w; x++){
1327                 int i= src[x + y*stride];
1328
1329                 if((unsigned)(i+thres1) > thres2){
1330                     if(i>=0){
1331                         i<<= QEXPSHIFT;
1332                         i/= qmul; //FIXME optimize
1333                         dst[x + y*stride]=  i;
1334                     }else{
1335                         i= -i;
1336                         i<<= QEXPSHIFT;
1337                         i/= qmul; //FIXME optimize
1338                         dst[x + y*stride]= -i;
1339                     }
1340                 }else
1341                     dst[x + y*stride]= 0;
1342             }
1343         }
1344     }else{
1345         for(y=0; y<h; y++){
1346             for(x=0; x<w; x++){
1347                 int i= src[x + y*stride];
1348
1349                 if((unsigned)(i+thres1) > thres2){
1350                     if(i>=0){
1351                         i<<= QEXPSHIFT;
1352                         i= (i + bias) / qmul; //FIXME optimize
1353                         dst[x + y*stride]=  i;
1354                     }else{
1355                         i= -i;
1356                         i<<= QEXPSHIFT;
1357                         i= (i + bias) / qmul; //FIXME optimize
1358                         dst[x + y*stride]= -i;
1359                     }
1360                 }else
1361                     dst[x + y*stride]= 0;
1362             }
1363         }
1364     }
1365 }
1366
1367 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1368     const int w= b->width;
1369     const int h= b->height;
1370     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1371     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1372     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1373     int x,y;
1374
1375     if(s->qlog == LOSSLESS_QLOG) return;
1376
1377     for(y=0; y<h; y++){
1378         for(x=0; x<w; x++){
1379             int i= src[x + y*stride];
1380             if(i<0){
1381                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1382             }else if(i>0){
1383                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1384             }
1385         }
1386     }
1387 }
1388
1389 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1390     const int w= b->width;
1391     const int h= b->height;
1392     int x,y;
1393
1394     for(y=h-1; y>=0; y--){
1395         for(x=w-1; x>=0; x--){
1396             int i= x + y*stride;
1397
1398             if(x){
1399                 if(use_median){
1400                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1401                     else  src[i] -= src[i - 1];
1402                 }else{
1403                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1404                     else  src[i] -= src[i - 1];
1405                 }
1406             }else{
1407                 if(y) src[i] -= src[i - stride];
1408             }
1409         }
1410     }
1411 }
1412
1413 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1414     const int w= b->width;
1415     const int h= b->height;
1416     int x,y;
1417
1418     for(y=0; y<h; y++){
1419         for(x=0; x<w; x++){
1420             int i= x + y*stride;
1421
1422             if(x){
1423                 if(use_median){
1424                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1425                     else  src[i] += src[i - 1];
1426                 }else{
1427                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1428                     else  src[i] += src[i - 1];
1429                 }
1430             }else{
1431                 if(y) src[i] += src[i - stride];
1432             }
1433         }
1434     }
1435 }
1436
1437 static void encode_qlogs(SnowContext *s){
1438     int plane_index, level, orientation;
1439
1440     for(plane_index=0; plane_index<2; plane_index++){
1441         for(level=0; level<s->spatial_decomposition_count; level++){
1442             for(orientation=level ? 1:0; orientation<4; orientation++){
1443                 if(orientation==2) continue;
1444                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1445             }
1446         }
1447     }
1448 }
1449
1450 static void encode_header(SnowContext *s){
1451     int plane_index, i;
1452     uint8_t kstate[32];
1453
1454     memset(kstate, MID_STATE, sizeof(kstate));
1455
1456     put_rac(&s->c, kstate, s->keyframe);
1457     if(s->keyframe || s->always_reset){
1458         ff_snow_reset_contexts(s);
1459         s->last_spatial_decomposition_type=
1460         s->last_qlog=
1461         s->last_qbias=
1462         s->last_mv_scale=
1463         s->last_block_max_depth= 0;
1464         for(plane_index=0; plane_index<2; plane_index++){
1465             Plane *p= &s->plane[plane_index];
1466             p->last_htaps=0;
1467             p->last_diag_mc=0;
1468             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1469         }
1470     }
1471     if(s->keyframe){
1472         put_symbol(&s->c, s->header_state, s->version, 0);
1473         put_rac(&s->c, s->header_state, s->always_reset);
1474         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1475         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1476         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1477         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1478         put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1479         put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1480         put_rac(&s->c, s->header_state, s->spatial_scalability);
1481 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1482         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1483
1484         encode_qlogs(s);
1485     }
1486
1487     if(!s->keyframe){
1488         int update_mc=0;
1489         for(plane_index=0; plane_index<2; plane_index++){
1490             Plane *p= &s->plane[plane_index];
1491             update_mc |= p->last_htaps   != p->htaps;
1492             update_mc |= p->last_diag_mc != p->diag_mc;
1493             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1494         }
1495         put_rac(&s->c, s->header_state, update_mc);
1496         if(update_mc){
1497             for(plane_index=0; plane_index<2; plane_index++){
1498                 Plane *p= &s->plane[plane_index];
1499                 put_rac(&s->c, s->header_state, p->diag_mc);
1500                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1501                 for(i= p->htaps/2; i; i--)
1502                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1503             }
1504         }
1505         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1506             put_rac(&s->c, s->header_state, 1);
1507             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1508             encode_qlogs(s);
1509         }else
1510             put_rac(&s->c, s->header_state, 0);
1511     }
1512
1513     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1514     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1515     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1516     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1517     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1518
1519 }
1520
1521 static void update_last_header_values(SnowContext *s){
1522     int plane_index;
1523
1524     if(!s->keyframe){
1525         for(plane_index=0; plane_index<2; plane_index++){
1526             Plane *p= &s->plane[plane_index];
1527             p->last_diag_mc= p->diag_mc;
1528             p->last_htaps  = p->htaps;
1529             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1530         }
1531     }
1532
1533     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1534     s->last_qlog                        = s->qlog;
1535     s->last_qbias                       = s->qbias;
1536     s->last_mv_scale                    = s->mv_scale;
1537     s->last_block_max_depth             = s->block_max_depth;
1538     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1539 }
1540
1541 static int qscale2qlog(int qscale){
1542     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1543            + 61*QROOT/8; ///< 64 > 60
1544 }
1545
1546 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1547 {
1548     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1549      * FIXME we know exact mv bits at this point,
1550      * but ratecontrol isn't set up to include them. */
1551     uint32_t coef_sum= 0;
1552     int level, orientation, delta_qlog;
1553
1554     for(level=0; level<s->spatial_decomposition_count; level++){
1555         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1556             SubBand *b= &s->plane[0].band[level][orientation];
1557             IDWTELEM *buf= b->ibuf;
1558             const int w= b->width;
1559             const int h= b->height;
1560             const int stride= b->stride;
1561             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1562             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1563             const int qdiv= (1<<16)/qmul;
1564             int x, y;
1565             //FIXME this is ugly
1566             for(y=0; y<h; y++)
1567                 for(x=0; x<w; x++)
1568                     buf[x+y*stride]= b->buf[x+y*stride];
1569             if(orientation==0)
1570                 decorrelate(s, b, buf, stride, 1, 0);
1571             for(y=0; y<h; y++)
1572                 for(x=0; x<w; x++)
1573                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1574         }
1575     }
1576
1577     /* ugly, ratecontrol just takes a sqrt again */
1578     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1579     assert(coef_sum < INT_MAX);
1580
1581     if(pict->pict_type == AV_PICTURE_TYPE_I){
1582         s->m.current_picture.mb_var_sum= coef_sum;
1583         s->m.current_picture.mc_mb_var_sum= 0;
1584     }else{
1585         s->m.current_picture.mc_mb_var_sum= coef_sum;
1586         s->m.current_picture.mb_var_sum= 0;
1587     }
1588
1589     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1590     if (pict->quality < 0)
1591         return INT_MIN;
1592     s->lambda= pict->quality * 3/2;
1593     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1594     s->qlog+= delta_qlog;
1595     return delta_qlog;
1596 }
1597
1598 static void calculate_visual_weight(SnowContext *s, Plane *p){
1599     int width = p->width;
1600     int height= p->height;
1601     int level, orientation, x, y;
1602
1603     for(level=0; level<s->spatial_decomposition_count; level++){
1604         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1605             SubBand *b= &p->band[level][orientation];
1606             IDWTELEM *ibuf= b->ibuf;
1607             int64_t error=0;
1608
1609             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1610             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1611             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1612             for(y=0; y<height; y++){
1613                 for(x=0; x<width; x++){
1614                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1615                     error += d*d;
1616                 }
1617             }
1618
1619             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1620         }
1621     }
1622 }
1623
1624 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1625                         const AVFrame *pict, int *got_packet)
1626 {
1627     SnowContext *s = avctx->priv_data;
1628     RangeCoder * const c= &s->c;
1629     AVFrame *pic = &s->new_picture;
1630     const int width= s->avctx->width;
1631     const int height= s->avctx->height;
1632     int level, orientation, plane_index, i, y, ret;
1633     uint8_t rc_header_bak[sizeof(s->header_state)];
1634     uint8_t rc_block_bak[sizeof(s->block_state)];
1635
1636     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1637         return ret;
1638
1639     ff_init_range_encoder(c, pkt->data, pkt->size);
1640     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1641
1642     for(i=0; i<3; i++){
1643         int hshift= i ? s->chroma_h_shift : 0;
1644         int vshift= i ? s->chroma_v_shift : 0;
1645         for(y=0; y<(height>>vshift); y++)
1646             memcpy(&s->input_picture.data[i][y * s->input_picture.linesize[i]],
1647                    &pict->data[i][y * pict->linesize[i]],
1648                    width>>hshift);
1649     }
1650     s->new_picture = *pict;
1651
1652     s->m.picture_number= avctx->frame_number;
1653     if(avctx->flags&CODEC_FLAG_PASS2){
1654         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1655         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1656         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1657             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1658             if (pic->quality < 0)
1659                 return -1;
1660         }
1661     }else{
1662         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1663         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1664     }
1665
1666     if(s->pass1_rc && avctx->frame_number == 0)
1667         pic->quality = 2*FF_QP2LAMBDA;
1668     if (pic->quality) {
1669         s->qlog   = qscale2qlog(pic->quality);
1670         s->lambda = pic->quality * 3/2;
1671     }
1672     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1673         s->qlog= LOSSLESS_QLOG;
1674         s->lambda = 0;
1675     }//else keep previous frame's qlog until after motion estimation
1676
1677     ff_snow_frame_start(s);
1678
1679     s->m.current_picture_ptr= &s->m.current_picture;
1680     s->m.last_picture.f.pts = s->m.current_picture.f.pts;
1681     s->m.current_picture.f.pts = pict->pts;
1682     if(pic->pict_type == AV_PICTURE_TYPE_P){
1683         int block_width = (width +15)>>4;
1684         int block_height= (height+15)>>4;
1685         int stride= s->current_picture.linesize[0];
1686
1687         assert(s->current_picture.data[0]);
1688         assert(s->last_picture[0].data[0]);
1689
1690         s->m.avctx= s->avctx;
1691         s->m.current_picture.f.data[0] = s->current_picture.data[0];
1692         s->m.   last_picture.f.data[0] = s->last_picture[0].data[0];
1693         s->m.    new_picture.f.data[0] = s->  input_picture.data[0];
1694         s->m.   last_picture_ptr= &s->m.   last_picture;
1695         s->m.linesize=
1696         s->m.   last_picture.f.linesize[0] =
1697         s->m.    new_picture.f.linesize[0] =
1698         s->m.current_picture.f.linesize[0] = stride;
1699         s->m.uvlinesize= s->current_picture.linesize[1];
1700         s->m.width = width;
1701         s->m.height= height;
1702         s->m.mb_width = block_width;
1703         s->m.mb_height= block_height;
1704         s->m.mb_stride=   s->m.mb_width+1;
1705         s->m.b8_stride= 2*s->m.mb_width+1;
1706         s->m.f_code=1;
1707         s->m.pict_type = pic->pict_type;
1708         s->m.me_method= s->avctx->me_method;
1709         s->m.me.scene_change_score=0;
1710         s->m.flags= s->avctx->flags;
1711         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1712         s->m.out_format= FMT_H263;
1713         s->m.unrestricted_mv= 1;
1714
1715         s->m.lambda = s->lambda;
1716         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1717         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1718
1719         s->m.dsp= s->dsp; //move
1720         ff_init_me(&s->m);
1721         s->dsp= s->m.dsp;
1722     }
1723
1724     if(s->pass1_rc){
1725         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1726         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1727     }
1728
1729 redo_frame:
1730
1731     if (pic->pict_type == AV_PICTURE_TYPE_I)
1732         s->spatial_decomposition_count= 5;
1733     else
1734         s->spatial_decomposition_count= 5;
1735
1736     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1737           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1738         s->spatial_decomposition_count--;
1739
1740     s->m.pict_type = pic->pict_type;
1741     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1742
1743     ff_snow_common_init_after_header(avctx);
1744
1745     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1746         for(plane_index=0; plane_index<3; plane_index++){
1747             calculate_visual_weight(s, &s->plane[plane_index]);
1748         }
1749     }
1750
1751     encode_header(s);
1752     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1753     encode_blocks(s, 1);
1754     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1755
1756     for(plane_index=0; plane_index<3; plane_index++){
1757         Plane *p= &s->plane[plane_index];
1758         int w= p->width;
1759         int h= p->height;
1760         int x, y;
1761 //        int bits= put_bits_count(&s->c.pb);
1762
1763         if (!s->memc_only) {
1764             //FIXME optimize
1765             if(pict->data[plane_index]) //FIXME gray hack
1766                 for(y=0; y<h; y++){
1767                     for(x=0; x<w; x++){
1768                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1769                     }
1770                 }
1771             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1772
1773             if(   plane_index==0
1774                && pic->pict_type == AV_PICTURE_TYPE_P
1775                && !(avctx->flags&CODEC_FLAG_PASS2)
1776                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1777                 ff_init_range_encoder(c, pkt->data, pkt->size);
1778                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1779                 pic->pict_type= AV_PICTURE_TYPE_I;
1780                 s->keyframe=1;
1781                 s->current_picture.key_frame=1;
1782                 goto redo_frame;
1783             }
1784
1785             if(s->qlog == LOSSLESS_QLOG){
1786                 for(y=0; y<h; y++){
1787                     for(x=0; x<w; x++){
1788                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1789                     }
1790                 }
1791             }else{
1792                 for(y=0; y<h; y++){
1793                     for(x=0; x<w; x++){
1794                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1795                     }
1796                 }
1797             }
1798
1799             /*  if(QUANTIZE2)
1800                 dwt_quantize(s, p, s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type);
1801             else*/
1802                 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1803
1804             if(s->pass1_rc && plane_index==0){
1805                 int delta_qlog = ratecontrol_1pass(s, pic);
1806                 if (delta_qlog <= INT_MIN)
1807                     return -1;
1808                 if(delta_qlog){
1809                     //reordering qlog in the bitstream would eliminate this reset
1810                     ff_init_range_encoder(c, pkt->data, pkt->size);
1811                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1812                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1813                     encode_header(s);
1814                     encode_blocks(s, 0);
1815                 }
1816             }
1817
1818             for(level=0; level<s->spatial_decomposition_count; level++){
1819                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1820                     SubBand *b= &p->band[level][orientation];
1821
1822                     if(!QUANTIZE2)
1823                         quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1824                     if(orientation==0)
1825                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1826                     if (!s->no_bitstream)
1827                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1828                     assert(b->parent==NULL || b->parent->stride == b->stride*2);
1829                     if(orientation==0)
1830                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1831                 }
1832             }
1833
1834             for(level=0; level<s->spatial_decomposition_count; level++){
1835                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1836                     SubBand *b= &p->band[level][orientation];
1837
1838                     dequantize(s, b, b->ibuf, b->stride);
1839                 }
1840             }
1841
1842             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1843             if(s->qlog == LOSSLESS_QLOG){
1844                 for(y=0; y<h; y++){
1845                     for(x=0; x<w; x++){
1846                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1847                     }
1848                 }
1849             }
1850             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1851         }else{
1852             //ME/MC only
1853             if(pic->pict_type == AV_PICTURE_TYPE_I){
1854                 for(y=0; y<h; y++){
1855                     for(x=0; x<w; x++){
1856                         s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
1857                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1858                     }
1859                 }
1860             }else{
1861                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1862                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1863             }
1864         }
1865         if(s->avctx->flags&CODEC_FLAG_PSNR){
1866             int64_t error= 0;
1867
1868             if(pict->data[plane_index]) //FIXME gray hack
1869                 for(y=0; y<h; y++){
1870                     for(x=0; x<w; x++){
1871                         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];
1872                         error += d*d;
1873                     }
1874                 }
1875             s->avctx->error[plane_index] += error;
1876             s->current_picture.error[plane_index] = error;
1877         }
1878
1879     }
1880
1881     update_last_header_values(s);
1882
1883     ff_snow_release_buffer(avctx);
1884
1885     s->current_picture.coded_picture_number = avctx->frame_number;
1886     s->current_picture.pict_type = pict->pict_type;
1887     s->current_picture.quality = pict->quality;
1888     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1889     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1890     s->m.current_picture.f.display_picture_number =
1891     s->m.current_picture.f.coded_picture_number   = avctx->frame_number;
1892     s->m.current_picture.f.quality                = pic->quality;
1893     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1894     if(s->pass1_rc)
1895         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1896             return -1;
1897     if(avctx->flags&CODEC_FLAG_PASS1)
1898         ff_write_pass1_stats(&s->m);
1899     s->m.last_pict_type = s->m.pict_type;
1900     avctx->frame_bits = s->m.frame_bits;
1901     avctx->mv_bits = s->m.mv_bits;
1902     avctx->misc_bits = s->m.misc_bits;
1903     avctx->p_tex_bits = s->m.p_tex_bits;
1904
1905     emms_c();
1906
1907     pkt->size = ff_rac_terminate(c);
1908     if (avctx->coded_frame->key_frame)
1909         pkt->flags |= AV_PKT_FLAG_KEY;
1910     *got_packet = 1;
1911
1912     return 0;
1913 }
1914
1915 static av_cold int encode_end(AVCodecContext *avctx)
1916 {
1917     SnowContext *s = avctx->priv_data;
1918
1919     ff_snow_common_end(s);
1920     if (s->input_picture.data[0])
1921         avctx->release_buffer(avctx, &s->input_picture);
1922     av_free(avctx->stats_out);
1923
1924     return 0;
1925 }
1926
1927 #define OFFSET(x) offsetof(SnowContext, x)
1928 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1929 static const AVOption options[] = {
1930     { "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 },
1931     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1932     { NULL },
1933 };
1934
1935 static const AVClass snowenc_class = {
1936     .class_name = "snow encoder",
1937     .item_name  = av_default_item_name,
1938     .option     = options,
1939     .version    = LIBAVUTIL_VERSION_INT,
1940 };
1941
1942 AVCodec ff_snow_encoder = {
1943     .name           = "snow",
1944     .type           = AVMEDIA_TYPE_VIDEO,
1945     .id             = AV_CODEC_ID_SNOW,
1946     .priv_data_size = sizeof(SnowContext),
1947     .init           = encode_init,
1948     .encode2        = encode_frame,
1949     .close          = encode_end,
1950     .pix_fmts       = (const enum AVPixelFormat[]){
1951         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1952         AV_PIX_FMT_NONE
1953     },
1954     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1955     .priv_class     = &snowenc_class,
1956 };
1957
1958
1959 #ifdef TEST
1960 #undef malloc
1961 #undef free
1962 #undef printf
1963
1964 #include "libavutil/lfg.h"
1965 #include "libavutil/mathematics.h"
1966
1967 int main(void){
1968 #define width  256
1969 #define height 256
1970     int buffer[2][width*height];
1971     SnowContext s;
1972     int i;
1973     AVLFG prng;
1974     s.spatial_decomposition_count=6;
1975     s.spatial_decomposition_type=1;
1976
1977     s.temp_dwt_buffer  = av_mallocz(width * sizeof(DWTELEM));
1978     s.temp_idwt_buffer = av_mallocz(width * sizeof(IDWTELEM));
1979
1980     av_lfg_init(&prng, 1);
1981
1982     printf("testing 5/3 DWT\n");
1983     for(i=0; i<width*height; i++)
1984         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1985
1986     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1987     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1988
1989     for(i=0; i<width*height; i++)
1990         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1991
1992     printf("testing 9/7 DWT\n");
1993     s.spatial_decomposition_type=0;
1994     for(i=0; i<width*height; i++)
1995         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1996
1997     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1998     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1999
2000     for(i=0; i<width*height; i++)
2001         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
2002
2003     {
2004     int level, orientation, x, y;
2005     int64_t errors[8][4];
2006     int64_t g=0;
2007
2008         memset(errors, 0, sizeof(errors));
2009         s.spatial_decomposition_count=3;
2010         s.spatial_decomposition_type=0;
2011         for(level=0; level<s.spatial_decomposition_count; level++){
2012             for(orientation=level ? 1 : 0; orientation<4; orientation++){
2013                 int w= width  >> (s.spatial_decomposition_count-level);
2014                 int h= height >> (s.spatial_decomposition_count-level);
2015                 int stride= width  << (s.spatial_decomposition_count-level);
2016                 DWTELEM *buf= buffer[0];
2017                 int64_t error=0;
2018
2019                 if(orientation&1) buf+=w;
2020                 if(orientation>1) buf+=stride>>1;
2021
2022                 memset(buffer[0], 0, sizeof(int)*width*height);
2023                 buf[w/2 + h/2*stride]= 256*256;
2024                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2025                 for(y=0; y<height; y++){
2026                     for(x=0; x<width; x++){
2027                         int64_t d= buffer[0][x + y*width];
2028                         error += d*d;
2029                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
2030                     }
2031                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
2032                 }
2033                 error= (int)(sqrt(error)+0.5);
2034                 errors[level][orientation]= error;
2035                 if(g) g=av_gcd(g, error);
2036                 else g= error;
2037             }
2038         }
2039         printf("static int const visual_weight[][4]={\n");
2040         for(level=0; level<s.spatial_decomposition_count; level++){
2041             printf("  {");
2042             for(orientation=0; orientation<4; orientation++){
2043                 printf("%8"PRId64",", errors[level][orientation]/g);
2044             }
2045             printf("},\n");
2046         }
2047         printf("};\n");
2048         {
2049             int level=2;
2050             int w= width  >> (s.spatial_decomposition_count-level);
2051             //int h= height >> (s.spatial_decomposition_count-level);
2052             int stride= width  << (s.spatial_decomposition_count-level);
2053             DWTELEM *buf= buffer[0];
2054             int64_t error=0;
2055
2056             buf+=w;
2057             buf+=stride>>1;
2058
2059             memset(buffer[0], 0, sizeof(int)*width*height);
2060             for(y=0; y<height; y++){
2061                 for(x=0; x<width; x++){
2062                     int tab[4]={0,2,3,1};
2063                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2064                 }
2065             }
2066             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2067             for(y=0; y<height; y++){
2068                 for(x=0; x<width; x++){
2069                     int64_t d= buffer[0][x + y*width];
2070                     error += d*d;
2071                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2072                 }
2073                 if(FFABS(height/2-y)<9) printf("\n");
2074             }
2075         }
2076
2077     }
2078     return 0;
2079 }
2080 #endif /* TEST */