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 Libav.
   5  *
   6  * Libav 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  * Libav 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 Libav; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 #include "libavutil/intmath.h"
  22 #include "libavutil/log.h"
  23 #include "libavutil/opt.h"
  24 #include "avcodec.h"
  25 #include "dsputil.h"
  26 #include "dwt.h"
  27 #include "snow.h"
  28
  29 #include "rangecoder.h"
  30 #include "mathops.h"
  31
  32 #include "mpegvideo.h"
  33 #include "h263.h"
  34
  35 #undef NDEBUG
  36 #include <assert.h>
  37
  38 #define QUANTIZE2 0
  39
  40 #if QUANTIZE2==1
  41 #define Q2_STEP 8
  42
  43 static void find_sse(SnowContext *s, Plane *p, int *score, int score_stride, IDWTELEM *r0, IDWTELEM *r1, int level, int orientation){
  44     SubBand *b= &p->band[level][orientation];
  45     int x, y;
  46     int xo=0;
  47     int yo=0;
  48     int step= 1 << (s->spatial_decomposition_count - level);
  49
  50     if(orientation&1)
  51         xo= step>>1;
  52     if(orientation&2)
  53         yo= step>>1;
  54
  55     //FIXME bias for nonzero ?
  56     //FIXME optimize
  57     memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
  58     for(y=0; y<p->height; y++){
  59         for(x=0; x<p->width; x++){
  60             int sx= (x-xo + step/2) / step / Q2_STEP;
  61             int sy= (y-yo + step/2) / step / Q2_STEP;
  62             int v= r0[x + y*p->width] - r1[x + y*p->width];
  63             assert(sx>=0 && sy>=0 && sx < score_stride);
  64             v= ((v+8)>>4)<<4;
  65             score[sx + sy*score_stride] += v*v;
  66             assert(score[sx + sy*score_stride] >= 0);
  67         }
  68     }
  69 }
  70
  71 static void dequantize_all(SnowContext *s, Plane *p, IDWTELEM *buffer, int width, int height){
  72     int level, orientation;
  73
  74     for(level=0; level<s->spatial_decomposition_count; level++){
  75         for(orientation=level ? 1 : 0; orientation<4; orientation++){
  76             SubBand *b= &p->band[level][orientation];
  77             IDWTELEM *dst= buffer + (b->ibuf - s->spatial_idwt_buffer);
  78
  79             dequantize(s, b, dst, b->stride);
  80         }
  81     }
  82 }
  83
  84 static void dwt_quantize(SnowContext *s, Plane *p, DWTELEM *buffer, int width, int height, int stride, int type){
  85     int level, orientation, ys, xs, x, y, pass;
  86     IDWTELEM best_dequant[height * stride];
  87     IDWTELEM idwt2_buffer[height * stride];
  88     const int score_stride= (width + 10)/Q2_STEP;
  89     int best_score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
  90     int score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
  91     int threshold= (s->m.lambda * s->m.lambda) >> 6;
  92
  93     //FIXME pass the copy cleanly ?
  94
  95 //    memcpy(dwt_buffer, buffer, height * stride * sizeof(DWTELEM));
  96     ff_spatial_dwt(buffer, s->temp_dwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
  97
  98     for(level=0; level<s->spatial_decomposition_count; level++){
  99         for(orientation=level ? 1 : 0; orientation<4; orientation++){
 100             SubBand *b= &p->band[level][orientation];
 101             IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
 102              DWTELEM *src=       buffer + (b-> buf - s->spatial_dwt_buffer);
 103             assert(src == b->buf); // code does not depend on this but it is true currently
 104
 105             quantize(s, b, dst, src, b->stride, s->qbias);
 106         }
 107     }
 108     for(pass=0; pass<1; pass++){
 109         if(s->qbias == 0) //keyframe
 110             continue;
 111         for(level=0; level<s->spatial_decomposition_count; level++){
 112             for(orientation=level ? 1 : 0; orientation<4; orientation++){
 113                 SubBand *b= &p->band[level][orientation];
 114                 IDWTELEM *dst= idwt2_buffer + (b->ibuf - s->spatial_idwt_buffer);
 115                 IDWTELEM *best_dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
 116
 117                 for(ys= 0; ys<Q2_STEP; ys++){
 118                     for(xs= 0; xs<Q2_STEP; xs++){
 119                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
 120                         dequantize_all(s, p, idwt2_buffer, width, height);
 121                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
 122                         find_sse(s, p, best_score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
 123                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
 124                         for(y=ys; y<b->height; y+= Q2_STEP){
 125                             for(x=xs; x<b->width; x+= Q2_STEP){
 126                                 if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
 127                                 if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
 128                                 //FIXME try more than just --
 129                             }
 130                         }
 131                         dequantize_all(s, p, idwt2_buffer, width, height);
 132                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
 133                         find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
 134                         for(y=ys; y<b->height; y+= Q2_STEP){
 135                             for(x=xs; x<b->width; x+= Q2_STEP){
 136                                 int score_idx= x/Q2_STEP + (y/Q2_STEP)*score_stride;
 137                                 if(score[score_idx] <= best_score[score_idx] + threshold){
 138                                     best_score[score_idx]= score[score_idx];
 139                                     if(best_dst[x + y*b->stride]<0) best_dst[x + y*b->stride]++;
 140                                     if(best_dst[x + y*b->stride]>0) best_dst[x + y*b->stride]--;
 141                                     //FIXME copy instead
 142                                 }
 143                             }
 144                         }
 145                     }
 146                 }
 147             }
 148         }
 149     }
 150     memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly instead of copy at the end
 151 }
 152
 153 #endif /* QUANTIZE2==1 */
 154
 155 static av_cold int encode_init(AVCodecContext *avctx)
 156 {
 157     SnowContext *s = avctx->priv_data;
 158     int plane_index, ret;
 159
 160     if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){
 161         av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"
 162                "Use vstrict=-2 / -strict -2 to use it anyway.\n");
 163         return -1;
 164     }
 165
 166     if(avctx->prediction_method == DWT_97
 167        && (avctx->flags & CODEC_FLAG_QSCALE)
 168        && avctx->global_quality == 0){
 169         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
 170         return -1;
 171     }
 172
 173     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
 174
 175     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
 176     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
 177
 178     for(plane_index=0; plane_index<3; plane_index++){
 179         s->plane[plane_index].diag_mc= 1;
 180         s->plane[plane_index].htaps= 6;
 181         s->plane[plane_index].hcoeff[0]=  40;
 182         s->plane[plane_index].hcoeff[1]= -10;
 183         s->plane[plane_index].hcoeff[2]=   2;
 184         s->plane[plane_index].fast_mc= 1;
 185     }
 186
 187     if ((ret = ff_snow_common_init(avctx)) < 0) {
 188         ff_snow_common_end(avctx->priv_data);
 189         return ret;
 190     }
 191     ff_snow_alloc_blocks(s);
 192
 193     s->version=0;
 194
 195     s->m.avctx   = avctx;
 196     s->m.flags   = avctx->flags;
 197     s->m.bit_rate= avctx->bit_rate;
 198
 199     s->m.me.temp      =
 200     s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
 201     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
 202     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
 203     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
 204     ff_h263_encode_init(&s->m); //mv_penalty
 205
 206     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
 207
 208     if(avctx->flags&CODEC_FLAG_PASS1){
 209         if(!avctx->stats_out)
 210             avctx->stats_out = av_mallocz(256);
 211     }
 212     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
 213         if(ff_rate_control_init(&s->m) < 0)
 214             return -1;
 215     }
 216     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
 217
 218     avctx->coded_frame= &s->current_picture;
 219     switch(avctx->pix_fmt){
 220 //    case PIX_FMT_YUV444P:
 221 //    case PIX_FMT_YUV422P:
 222     case PIX_FMT_YUV420P:
 223     case PIX_FMT_GRAY8:
 224 //    case PIX_FMT_YUV411P:
 225 //    case PIX_FMT_YUV410P:
 226         s->colorspace_type= 0;
 227         break;
 228 /*    case PIX_FMT_RGB32:
 229         s->colorspace= 1;
 230         break;*/
 231     default:
 232         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
 233         return -1;
 234     }
 235 //    avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
 236     s->chroma_h_shift= 1;
 237     s->chroma_v_shift= 1;
 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     s->avctx->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)
 258 {
 259     int s, i, j;
 260
 261     s = 0;
 262     for (i = 0; i < w; 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 + y*uvstride)*block_w/2,
 329                                 s->input_picture.data[2] + (x + y*uvstride)*block_w/2};
 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);
 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>>2;
 454     sum = pix_sum(current_data[1], uvstride, block_w>>1);
 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>>1);
 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/2 : block_size;
 580     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+1] : ff_obmc_tab[s->block_max_depth];
 581     const int obmc_stride= plane_index ? block_size : 2*block_size;
 582     const int ref_stride= s->current_picture.linesize[plane_index];
 583     uint8_t *src= s-> input_picture.data[plane_index];
 584     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
 585     const int b_stride = s->b_width << s->block_max_depth;
 586     const int w= p->width;
 587     const int h= p->height;
 588     int index= mb_x + mb_y*b_stride;
 589     BlockNode *b= &s->block[index];
 590     BlockNode backup= *b;
 591     int ab=0;
 592     int aa=0;
 593
 594     b->type|= BLOCK_INTRA;
 595     b->color[plane_index]= 0;
 596     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
 597
 598     for(i=0; i<4; i++){
 599         int mb_x2= mb_x + (i &1) - 1;
 600         int mb_y2= mb_y + (i>>1) - 1;
 601         int x= block_w*mb_x2 + block_w/2;
 602         int y= block_w*mb_y2 + block_w/2;
 603
 604         add_yblock(s, 0, NULL, dst + ((i&1)+(i>>1)*obmc_stride)*block_w, NULL, obmc,
 605                     x, y, block_w, block_w, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
 606
 607         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_w); y2++){
 608             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
 609                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_w*mb_y - block_w/2))*obmc_stride;
 610                 int obmc_v= obmc[index];
 611                 int d;
 612                 if(y<0) obmc_v += obmc[index + block_w*obmc_stride];
 613                 if(x<0) obmc_v += obmc[index + block_w];
 614                 if(y+block_w>h) obmc_v += obmc[index - block_w*obmc_stride];
 615                 if(x+block_w>w) obmc_v += obmc[index - block_w];
 616                 //FIXME precalculate this or simplify it somehow else
 617
 618                 d = -dst[index] + (1<<(FRAC_BITS-1));
 619                 dst[index] = d;
 620                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
 621                 aa += obmc_v * obmc_v; //FIXME precalculate this
 622             }
 623         }
 624     }
 625     *b= backup;
 626
 627     return av_clip(((ab<<LOG2_OBMC_MAX) + aa/2)/aa, 0, 255); //FIXME we should not need clipping
 628 }
 629
 630 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
 631     const int b_stride = s->b_width << s->block_max_depth;
 632     const int b_height = s->b_height<< s->block_max_depth;
 633     int index= x + y*b_stride;
 634     const BlockNode *b     = &s->block[index];
 635     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
 636     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
 637     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
 638     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
 639     int dmx, dmy;
 640 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
 641 //  int my_context= av_log2(2*FFABS(left->my - top->my));
 642
 643     if(x<0 || x>=b_stride || y>=b_height)
 644         return 0;
 645 /*
 646 1            0      0
 647 01X          1-2    1
 648 001XX        3-6    2-3
 649 0001XXX      7-14   4-7
 650 00001XXXX   15-30   8-15
 651 */
 652 //FIXME try accurate rate
 653 //FIXME intra and inter predictors if surrounding blocks are not the same type
 654     if(b->type & BLOCK_INTRA){
 655         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
 656                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
 657                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
 658     }else{
 659         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
 660         dmx-= b->mx;
 661         dmy-= b->my;
 662         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
 663                     + av_log2(2*FFABS(dmy))
 664                     + av_log2(2*b->ref));
 665     }
 666 }
 667
 668 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
 669     Plane *p= &s->plane[plane_index];
 670     const int block_size = MB_SIZE >> s->block_max_depth;
 671     const int block_w    = plane_index ? block_size/2 : block_size;
 672     const int obmc_stride= plane_index ? block_size : 2*block_size;
 673     const int ref_stride= s->current_picture.linesize[plane_index];
 674     uint8_t *dst= s->current_picture.data[plane_index];
 675     uint8_t *src= s->  input_picture.data[plane_index];
 676     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
 677     uint8_t *cur = s->scratchbuf;
 678     uint8_t *tmp = s->emu_edge_buffer;
 679     const int b_stride = s->b_width << s->block_max_depth;
 680     const int b_height = s->b_height<< s->block_max_depth;
 681     const int w= p->width;
 682     const int h= p->height;
 683     int distortion;
 684     int rate= 0;
 685     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 686     int sx= block_w*mb_x - block_w/2;
 687     int sy= block_w*mb_y - block_w/2;
 688     int x0= FFMAX(0,-sx);
 689     int y0= FFMAX(0,-sy);
 690     int x1= FFMIN(block_w*2, w-sx);
 691     int y1= FFMIN(block_w*2, h-sy);
 692     int i,x,y;
 693
 694     ff_snow_pred_block(s, cur, tmp, ref_stride, sx, sy, block_w*2, block_w*2, &s->block[mb_x + mb_y*b_stride], plane_index, w, h);
 695
 696     for(y=y0; y<y1; y++){
 697         const uint8_t *obmc1= obmc_edged[y];
 698         const IDWTELEM *pred1 = pred + y*obmc_stride;
 699         uint8_t *cur1 = cur + y*ref_stride;
 700         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
 701         for(x=x0; x<x1; x++){
 702 #if FRAC_BITS >= LOG2_OBMC_MAX
 703             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
 704 #else
 705             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
 706 #endif
 707             v = (v + pred1[x]) >> FRAC_BITS;
 708             if(v&(~255)) v= ~(v>>31);
 709             dst1[x] = v;
 710         }
 711     }
 712
 713     /* copy the regions where obmc[] = (uint8_t)256 */
 714     if(LOG2_OBMC_MAX == 8
 715         && (mb_x == 0 || mb_x == b_stride-1)
 716         && (mb_y == 0 || mb_y == b_height-1)){
 717         if(mb_x == 0)
 718             x1 = block_w;
 719         else
 720             x0 = block_w;
 721         if(mb_y == 0)
 722             y1 = block_w;
 723         else
 724             y0 = block_w;
 725         for(y=y0; y<y1; y++)
 726             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
 727     }
 728
 729     if(block_w==16){
 730         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
 731         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
 732         /* FIXME cmps overlap but do not cover the wavelet's whole support.
 733          * So improving the score of one block is not strictly guaranteed
 734          * to improve the score of the whole frame, thus iterative motion
 735          * estimation does not always converge. */
 736         if(s->avctx->me_cmp == FF_CMP_W97)
 737             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 738         else if(s->avctx->me_cmp == FF_CMP_W53)
 739             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
 740         else{
 741             distortion = 0;
 742             for(i=0; i<4; i++){
 743                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
 744                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
 745             }
 746         }
 747     }else{
 748         assert(block_w==8);
 749         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
 750     }
 751
 752     if(plane_index==0){
 753         for(i=0; i<4; i++){
 754 /* ..RRr
 755  * .RXx.
 756  * rxx..
 757  */
 758             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
 759         }
 760         if(mb_x == b_stride-2)
 761             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
 762     }
 763     return distortion + rate*penalty_factor;
 764 }
 765
 766 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
 767     int i, y2;
 768     Plane *p= &s->plane[plane_index];
 769     const int block_size = MB_SIZE >> s->block_max_depth;
 770     const int block_w    = plane_index ? block_size/2 : block_size;
 771     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+1] : ff_obmc_tab[s->block_max_depth];
 772     const int obmc_stride= plane_index ? block_size : 2*block_size;
 773     const int ref_stride= s->current_picture.linesize[plane_index];
 774     uint8_t *dst= s->current_picture.data[plane_index];
 775     uint8_t *src= s-> input_picture.data[plane_index];
 776     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
 777     // const has only been removed from zero_dst to suppress a warning
 778     static IDWTELEM zero_dst[4096]; //FIXME
 779     const int b_stride = s->b_width << s->block_max_depth;
 780     const int w= p->width;
 781     const int h= p->height;
 782     int distortion= 0;
 783     int rate= 0;
 784     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
 785
 786     for(i=0; i<9; i++){
 787         int mb_x2= mb_x + (i%3) - 1;
 788         int mb_y2= mb_y + (i/3) - 1;
 789         int x= block_w*mb_x2 + block_w/2;
 790         int y= block_w*mb_y2 + block_w/2;
 791
 792         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
 793                    x, y, block_w, block_w, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
 794
 795         //FIXME find a cleaner/simpler way to skip the outside stuff
 796         for(y2= y; y2<0; y2++)
 797             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 798         for(y2= h; y2<y+block_w; y2++)
 799             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
 800         if(x<0){
 801             for(y2= y; y2<y+block_w; y2++)
 802                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
 803         }
 804         if(x+block_w > w){
 805             for(y2= y; y2<y+block_w; y2++)
 806                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
 807         }
 808
 809         assert(block_w== 8 || block_w==16);
 810         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_w);
 811     }
 812
 813     if(plane_index==0){
 814         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
 815         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
 816
 817 /* ..RRRr
 818  * .RXXx.
 819  * .RXXx.
 820  * rxxx.
 821  */
 822         if(merged)
 823             rate = get_block_bits(s, mb_x, mb_y, 2);
 824         for(i=merged?4:0; i<9; i++){
 825             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
 826             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
 827         }
 828     }
 829     return distortion + rate*penalty_factor;
 830 }
 831
 832 static int encode_subband_c0run(SnowContext *s, SubBand *b, IDWTELEM *src, IDWTELEM *parent, int stride, int orientation){
 833     const int w= b->width;
 834     const int h= b->height;
 835     int x, y;
 836
 837     if(1){
 838         int run=0;
 839         int *runs = s->run_buffer;
 840         int run_index=0;
 841         int max_index;
 842
 843         for(y=0; y<h; y++){
 844             for(x=0; x<w; x++){
 845                 int v, p=0;
 846                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 847                 v= src[x + y*stride];
 848
 849                 if(y){
 850                     t= src[x + (y-1)*stride];
 851                     if(x){
 852                         lt= src[x - 1 + (y-1)*stride];
 853                     }
 854                     if(x + 1 < w){
 855                         rt= src[x + 1 + (y-1)*stride];
 856                     }
 857                 }
 858                 if(x){
 859                     l= src[x - 1 + y*stride];
 860                     /*if(x > 1){
 861                         if(orientation==1) ll= src[y + (x-2)*stride];
 862                         else               ll= src[x - 2 + y*stride];
 863                     }*/
 864                 }
 865                 if(parent){
 866                     int px= x>>1;
 867                     int py= y>>1;
 868                     if(px<b->parent->width && py<b->parent->height)
 869                         p= parent[px + py*2*stride];
 870                 }
 871                 if(!(/*ll|*/l|lt|t|rt|p)){
 872                     if(v){
 873                         runs[run_index++]= run;
 874                         run=0;
 875                     }else{
 876                         run++;
 877                     }
 878                 }
 879             }
 880         }
 881         max_index= run_index;
 882         runs[run_index++]= run;
 883         run_index=0;
 884         run= runs[run_index++];
 885
 886         put_symbol2(&s->c, b->state[30], max_index, 0);
 887         if(run_index <= max_index)
 888             put_symbol2(&s->c, b->state[1], run, 3);
 889
 890         for(y=0; y<h; y++){
 891             if(s->c.bytestream_end - s->c.bytestream < w*40){
 892                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 893                 return -1;
 894             }
 895             for(x=0; x<w; x++){
 896                 int v, p=0;
 897                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
 898                 v= src[x + y*stride];
 899
 900                 if(y){
 901                     t= src[x + (y-1)*stride];
 902                     if(x){
 903                         lt= src[x - 1 + (y-1)*stride];
 904                     }
 905                     if(x + 1 < w){
 906                         rt= src[x + 1 + (y-1)*stride];
 907                     }
 908                 }
 909                 if(x){
 910                     l= src[x - 1 + y*stride];
 911                     /*if(x > 1){
 912                         if(orientation==1) ll= src[y + (x-2)*stride];
 913                         else               ll= src[x - 2 + y*stride];
 914                     }*/
 915                 }
 916                 if(parent){
 917                     int px= x>>1;
 918                     int py= y>>1;
 919                     if(px<b->parent->width && py<b->parent->height)
 920                         p= parent[px + py*2*stride];
 921                 }
 922                 if(/*ll|*/l|lt|t|rt|p){
 923                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 924
 925                     put_rac(&s->c, &b->state[0][context], !!v);
 926                 }else{
 927                     if(!run){
 928                         run= runs[run_index++];
 929
 930                         if(run_index <= max_index)
 931                             put_symbol2(&s->c, b->state[1], run, 3);
 932                         assert(v);
 933                     }else{
 934                         run--;
 935                         assert(!v);
 936                     }
 937                 }
 938                 if(v){
 939                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
 940                     int l2= 2*FFABS(l) + (l<0);
 941                     int t2= 2*FFABS(t) + (t<0);
 942
 943                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
 944                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
 945                 }
 946             }
 947         }
 948     }
 949     return 0;
 950 }
 951
 952 static int encode_subband(SnowContext *s, SubBand *b, IDWTELEM *src, IDWTELEM *parent, int stride, int orientation){
 953 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
 954 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
 955     return encode_subband_c0run(s, b, src, parent, stride, orientation);
 956 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
 957 }
 958
 959 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){
 960     const int b_stride= s->b_width << s->block_max_depth;
 961     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
 962     BlockNode backup= *block;
 963     unsigned value;
 964     int rd, index;
 965
 966     assert(mb_x>=0 && mb_y>=0);
 967     assert(mb_x<b_stride);
 968
 969     if(intra){
 970         block->color[0] = p[0];
 971         block->color[1] = p[1];
 972         block->color[2] = p[2];
 973         block->type |= BLOCK_INTRA;
 974     }else{
 975         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
 976         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
 977         if(s->me_cache[index] == value)
 978             return 0;
 979         s->me_cache[index]= value;
 980
 981         block->mx= p[0];
 982         block->my= p[1];
 983         block->type &= ~BLOCK_INTRA;
 984     }
 985
 986     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
 987
 988 //FIXME chroma
 989     if(rd < *best_rd){
 990         *best_rd= rd;
 991         return 1;
 992     }else{
 993         *block= backup;
 994         return 0;
 995     }
 996 }
 997
 998 /* special case for int[2] args we discard afterwards,
 999  * fixes compilation problem with gcc 2.95 */
1000 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){
1001     int p[2] = {p0, p1};
1002     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
1003 }
1004
1005 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){
1006     const int b_stride= s->b_width << s->block_max_depth;
1007     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
1008     BlockNode backup[4];
1009     unsigned value;
1010     int rd, index;
1011
1012     /* We don't initialize backup[] during variable declaration, because
1013      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
1014      * 'int16_t'". */
1015     backup[0] = block[0];
1016     backup[1] = block[1];
1017     backup[2] = block[b_stride];
1018     backup[3] = block[b_stride + 1];
1019
1020     assert(mb_x>=0 && mb_y>=0);
1021     assert(mb_x<b_stride);
1022     assert(((mb_x|mb_y)&1) == 0);
1023
1024     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
1025     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
1026     if(s->me_cache[index] == value)
1027         return 0;
1028     s->me_cache[index]= value;
1029
1030     block->mx= p0;
1031     block->my= p1;
1032     block->ref= ref;
1033     block->type &= ~BLOCK_INTRA;
1034     block[1]= block[b_stride]= block[b_stride+1]= *block;
1035
1036     rd= get_4block_rd(s, mb_x, mb_y, 0);
1037
1038 //FIXME chroma
1039     if(rd < *best_rd){
1040         *best_rd= rd;
1041         return 1;
1042     }else{
1043         block[0]= backup[0];
1044         block[1]= backup[1];
1045         block[b_stride]= backup[2];
1046         block[b_stride+1]= backup[3];
1047         return 0;
1048     }
1049 }
1050
1051 static void iterative_me(SnowContext *s){
1052     int pass, mb_x, mb_y;
1053     const int b_width = s->b_width  << s->block_max_depth;
1054     const int b_height= s->b_height << s->block_max_depth;
1055     const int b_stride= b_width;
1056     int color[3];
1057
1058     {
1059         RangeCoder r = s->c;
1060         uint8_t state[sizeof(s->block_state)];
1061         memcpy(state, s->block_state, sizeof(s->block_state));
1062         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1063             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1064                 encode_q_branch(s, 0, mb_x, mb_y);
1065         s->c = r;
1066         memcpy(s->block_state, state, sizeof(s->block_state));
1067     }
1068
1069     for(pass=0; pass<25; pass++){
1070         int change= 0;
1071
1072         for(mb_y= 0; mb_y<b_height; mb_y++){
1073             for(mb_x= 0; mb_x<b_width; mb_x++){
1074                 int dia_change, i, j, ref;
1075                 int best_rd= INT_MAX, ref_rd;
1076                 BlockNode backup, ref_b;
1077                 const int index= mb_x + mb_y * b_stride;
1078                 BlockNode *block= &s->block[index];
1079                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1080                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1081                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1082                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1083                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1084                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1085                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1086                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1087                 const int b_w= (MB_SIZE >> s->block_max_depth);
1088                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1089
1090                 if(pass && (block->type & BLOCK_OPT))
1091                     continue;
1092                 block->type |= BLOCK_OPT;
1093
1094                 backup= *block;
1095
1096                 if(!s->me_cache_generation)
1097                     memset(s->me_cache, 0, sizeof(s->me_cache));
1098                 s->me_cache_generation += 1<<22;
1099
1100                 //FIXME precalculate
1101                 {
1102                     int x, y;
1103                     for (y = 0; y < b_w * 2; y++)
1104                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1105                     if(mb_x==0)
1106                         for(y=0; y<b_w*2; y++)
1107                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1108                     if(mb_x==b_stride-1)
1109                         for(y=0; y<b_w*2; y++)
1110                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1111                     if(mb_y==0){
1112                         for(x=0; x<b_w*2; x++)
1113                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1114                         for(y=1; y<b_w; y++)
1115                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1116                     }
1117                     if(mb_y==b_height-1){
1118                         for(x=0; x<b_w*2; x++)
1119                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1120                         for(y=b_w; y<b_w*2-1; y++)
1121                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1122                     }
1123                 }
1124
1125                 //skip stuff outside the picture
1126                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1127                     uint8_t *src= s->  input_picture.data[0];
1128                     uint8_t *dst= s->current_picture.data[0];
1129                     const int stride= s->current_picture.linesize[0];
1130                     const int block_w= MB_SIZE >> s->block_max_depth;
1131                     const int sx= block_w*mb_x - block_w/2;
1132                     const int sy= block_w*mb_y - block_w/2;
1133                     const int w= s->plane[0].width;
1134                     const int h= s->plane[0].height;
1135                     int y;
1136
1137                     for(y=sy; y<0; y++)
1138                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1139                     for(y=h; y<sy+block_w*2; y++)
1140                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1141                     if(sx<0){
1142                         for(y=sy; y<sy+block_w*2; y++)
1143                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1144                     }
1145                     if(sx+block_w*2 > w){
1146                         for(y=sy; y<sy+block_w*2; y++)
1147                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1148                     }
1149                 }
1150
1151                 // intra(black) = neighbors' contribution to the current block
1152                 for(i=0; i<3; i++)
1153                     color[i]= get_dc(s, mb_x, mb_y, i);
1154
1155                 // get previous score (cannot be cached due to OBMC)
1156                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1157                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1158                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1159                 }else
1160                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1161
1162                 ref_b= *block;
1163                 ref_rd= best_rd;
1164                 for(ref=0; ref < s->ref_frames; ref++){
1165                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1166                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1167                         continue;
1168                     block->ref= ref;
1169                     best_rd= INT_MAX;
1170
1171                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1172                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1173                     if(tb)
1174                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1175                     if(lb)
1176                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1177                     if(rb)
1178                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1179                     if(bb)
1180                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1181
1182                     /* fullpel ME */
1183                     //FIXME avoid subpel interpolation / round to nearest integer
1184                     do{
1185                         dia_change=0;
1186                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1187                             for(j=0; j<i; j++){
1188                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1189                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1190                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1191                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1192                             }
1193                         }
1194                     }while(dia_change);
1195                     /* subpel ME */
1196                     do{
1197                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1198                         dia_change=0;
1199                         for(i=0; i<8; i++)
1200                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1201                     }while(dia_change);
1202                     //FIXME or try the standard 2 pass qpel or similar
1203
1204                     mvr[0][0]= block->mx;
1205                     mvr[0][1]= block->my;
1206                     if(ref_rd > best_rd){
1207                         ref_rd= best_rd;
1208                         ref_b= *block;
1209                     }
1210                 }
1211                 best_rd= ref_rd;
1212                 *block= ref_b;
1213                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1214                 //FIXME RD style color selection
1215                 if(!same_block(block, &backup)){
1216                     if(tb ) tb ->type &= ~BLOCK_OPT;
1217                     if(lb ) lb ->type &= ~BLOCK_OPT;
1218                     if(rb ) rb ->type &= ~BLOCK_OPT;
1219                     if(bb ) bb ->type &= ~BLOCK_OPT;
1220                     if(tlb) tlb->type &= ~BLOCK_OPT;
1221                     if(trb) trb->type &= ~BLOCK_OPT;
1222                     if(blb) blb->type &= ~BLOCK_OPT;
1223                     if(brb) brb->type &= ~BLOCK_OPT;
1224                     change ++;
1225                 }
1226             }
1227         }
1228         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1229         if(!change)
1230             break;
1231     }
1232
1233     if(s->block_max_depth == 1){
1234         int change= 0;
1235         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1236             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1237                 int i;
1238                 int best_rd, init_rd;
1239                 const int index= mb_x + mb_y * b_stride;
1240                 BlockNode *b[4];
1241
1242                 b[0]= &s->block[index];
1243                 b[1]= b[0]+1;
1244                 b[2]= b[0]+b_stride;
1245                 b[3]= b[2]+1;
1246                 if(same_block(b[0], b[1]) &&
1247                    same_block(b[0], b[2]) &&
1248                    same_block(b[0], b[3]))
1249                     continue;
1250
1251                 if(!s->me_cache_generation)
1252                     memset(s->me_cache, 0, sizeof(s->me_cache));
1253                 s->me_cache_generation += 1<<22;
1254
1255                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1256
1257                 //FIXME more multiref search?
1258                 check_4block_inter(s, mb_x, mb_y,
1259                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1260                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1261
1262                 for(i=0; i<4; i++)
1263                     if(!(b[i]->type&BLOCK_INTRA))
1264                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1265
1266                 if(init_rd != best_rd)
1267                     change++;
1268             }
1269         }
1270         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1271     }
1272 }
1273
1274 static void encode_blocks(SnowContext *s, int search){
1275     int x, y;
1276     int w= s->b_width;
1277     int h= s->b_height;
1278
1279     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1280         iterative_me(s);
1281
1282     for(y=0; y<h; y++){
1283         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1284             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1285             return;
1286         }
1287         for(x=0; x<w; x++){
1288             if(s->avctx->me_method == ME_ITER || !search)
1289                 encode_q_branch2(s, 0, x, y);
1290             else
1291                 encode_q_branch (s, 0, x, y);
1292         }
1293     }
1294 }
1295
1296 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1297     const int w= b->width;
1298     const int h= b->height;
1299     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1300     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1301     int x,y, thres1, thres2;
1302
1303     if(s->qlog == LOSSLESS_QLOG){
1304         for(y=0; y<h; y++)
1305             for(x=0; x<w; x++)
1306                 dst[x + y*stride]= src[x + y*stride];
1307         return;
1308     }
1309
1310     bias= bias ? 0 : (3*qmul)>>3;
1311     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1312     thres2= 2*thres1;
1313
1314     if(!bias){
1315         for(y=0; y<h; y++){
1316             for(x=0; x<w; x++){
1317                 int i= src[x + y*stride];
1318
1319                 if((unsigned)(i+thres1) > thres2){
1320                     if(i>=0){
1321                         i<<= QEXPSHIFT;
1322                         i/= qmul; //FIXME optimize
1323                         dst[x + y*stride]=  i;
1324                     }else{
1325                         i= -i;
1326                         i<<= QEXPSHIFT;
1327                         i/= qmul; //FIXME optimize
1328                         dst[x + y*stride]= -i;
1329                     }
1330                 }else
1331                     dst[x + y*stride]= 0;
1332             }
1333         }
1334     }else{
1335         for(y=0; y<h; y++){
1336             for(x=0; x<w; x++){
1337                 int i= src[x + y*stride];
1338
1339                 if((unsigned)(i+thres1) > thres2){
1340                     if(i>=0){
1341                         i<<= QEXPSHIFT;
1342                         i= (i + bias) / qmul; //FIXME optimize
1343                         dst[x + y*stride]=  i;
1344                     }else{
1345                         i= -i;
1346                         i<<= QEXPSHIFT;
1347                         i= (i + bias) / qmul; //FIXME optimize
1348                         dst[x + y*stride]= -i;
1349                     }
1350                 }else
1351                     dst[x + y*stride]= 0;
1352             }
1353         }
1354     }
1355 }
1356
1357 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1358     const int w= b->width;
1359     const int h= b->height;
1360     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1361     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1362     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1363     int x,y;
1364
1365     if(s->qlog == LOSSLESS_QLOG) return;
1366
1367     for(y=0; y<h; y++){
1368         for(x=0; x<w; x++){
1369             int i= src[x + y*stride];
1370             if(i<0){
1371                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1372             }else if(i>0){
1373                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1374             }
1375         }
1376     }
1377 }
1378
1379 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1380     const int w= b->width;
1381     const int h= b->height;
1382     int x,y;
1383
1384     for(y=h-1; y>=0; y--){
1385         for(x=w-1; x>=0; x--){
1386             int i= x + y*stride;
1387
1388             if(x){
1389                 if(use_median){
1390                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1391                     else  src[i] -= src[i - 1];
1392                 }else{
1393                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1394                     else  src[i] -= src[i - 1];
1395                 }
1396             }else{
1397                 if(y) src[i] -= src[i - stride];
1398             }
1399         }
1400     }
1401 }
1402
1403 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1404     const int w= b->width;
1405     const int h= b->height;
1406     int x,y;
1407
1408     for(y=0; y<h; y++){
1409         for(x=0; x<w; x++){
1410             int i= x + y*stride;
1411
1412             if(x){
1413                 if(use_median){
1414                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1415                     else  src[i] += src[i - 1];
1416                 }else{
1417                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1418                     else  src[i] += src[i - 1];
1419                 }
1420             }else{
1421                 if(y) src[i] += src[i - stride];
1422             }
1423         }
1424     }
1425 }
1426
1427 static void encode_qlogs(SnowContext *s){
1428     int plane_index, level, orientation;
1429
1430     for(plane_index=0; plane_index<2; plane_index++){
1431         for(level=0; level<s->spatial_decomposition_count; level++){
1432             for(orientation=level ? 1:0; orientation<4; orientation++){
1433                 if(orientation==2) continue;
1434                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1435             }
1436         }
1437     }
1438 }
1439
1440 static void encode_header(SnowContext *s){
1441     int plane_index, i;
1442     uint8_t kstate[32];
1443
1444     memset(kstate, MID_STATE, sizeof(kstate));
1445
1446     put_rac(&s->c, kstate, s->keyframe);
1447     if(s->keyframe || s->always_reset){
1448         ff_snow_reset_contexts(s);
1449         s->last_spatial_decomposition_type=
1450         s->last_qlog=
1451         s->last_qbias=
1452         s->last_mv_scale=
1453         s->last_block_max_depth= 0;
1454         for(plane_index=0; plane_index<2; plane_index++){
1455             Plane *p= &s->plane[plane_index];
1456             p->last_htaps=0;
1457             p->last_diag_mc=0;
1458             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1459         }
1460     }
1461     if(s->keyframe){
1462         put_symbol(&s->c, s->header_state, s->version, 0);
1463         put_rac(&s->c, s->header_state, s->always_reset);
1464         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1465         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1466         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1467         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1468         put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1469         put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1470         put_rac(&s->c, s->header_state, s->spatial_scalability);
1471 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1472         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1473
1474         encode_qlogs(s);
1475     }
1476
1477     if(!s->keyframe){
1478         int update_mc=0;
1479         for(plane_index=0; plane_index<2; plane_index++){
1480             Plane *p= &s->plane[plane_index];
1481             update_mc |= p->last_htaps   != p->htaps;
1482             update_mc |= p->last_diag_mc != p->diag_mc;
1483             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1484         }
1485         put_rac(&s->c, s->header_state, update_mc);
1486         if(update_mc){
1487             for(plane_index=0; plane_index<2; plane_index++){
1488                 Plane *p= &s->plane[plane_index];
1489                 put_rac(&s->c, s->header_state, p->diag_mc);
1490                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1491                 for(i= p->htaps/2; i; i--)
1492                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1493             }
1494         }
1495         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1496             put_rac(&s->c, s->header_state, 1);
1497             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1498             encode_qlogs(s);
1499         }else
1500             put_rac(&s->c, s->header_state, 0);
1501     }
1502
1503     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1504     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1505     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1506     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1507     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1508
1509 }
1510
1511 static void update_last_header_values(SnowContext *s){
1512     int plane_index;
1513
1514     if(!s->keyframe){
1515         for(plane_index=0; plane_index<2; plane_index++){
1516             Plane *p= &s->plane[plane_index];
1517             p->last_diag_mc= p->diag_mc;
1518             p->last_htaps  = p->htaps;
1519             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1520         }
1521     }
1522
1523     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1524     s->last_qlog                        = s->qlog;
1525     s->last_qbias                       = s->qbias;
1526     s->last_mv_scale                    = s->mv_scale;
1527     s->last_block_max_depth             = s->block_max_depth;
1528     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1529 }
1530
1531 static int qscale2qlog(int qscale){
1532     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1533            + 61*QROOT/8; ///< 64 > 60
1534 }
1535
1536 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1537 {
1538     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1539      * FIXME we know exact mv bits at this point,
1540      * but ratecontrol isn't set up to include them. */
1541     uint32_t coef_sum= 0;
1542     int level, orientation, delta_qlog;
1543
1544     for(level=0; level<s->spatial_decomposition_count; level++){
1545         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1546             SubBand *b= &s->plane[0].band[level][orientation];
1547             IDWTELEM *buf= b->ibuf;
1548             const int w= b->width;
1549             const int h= b->height;
1550             const int stride= b->stride;
1551             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1552             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1553             const int qdiv= (1<<16)/qmul;
1554             int x, y;
1555             //FIXME this is ugly
1556             for(y=0; y<h; y++)
1557                 for(x=0; x<w; x++)
1558                     buf[x+y*stride]= b->buf[x+y*stride];
1559             if(orientation==0)
1560                 decorrelate(s, b, buf, stride, 1, 0);
1561             for(y=0; y<h; y++)
1562                 for(x=0; x<w; x++)
1563                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1564         }
1565     }
1566
1567     /* ugly, ratecontrol just takes a sqrt again */
1568     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1569     assert(coef_sum < INT_MAX);
1570
1571     if(pict->pict_type == AV_PICTURE_TYPE_I){
1572         s->m.current_picture.mb_var_sum= coef_sum;
1573         s->m.current_picture.mc_mb_var_sum= 0;
1574     }else{
1575         s->m.current_picture.mc_mb_var_sum= coef_sum;
1576         s->m.current_picture.mb_var_sum= 0;
1577     }
1578
1579     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1580     if (pict->quality < 0)
1581         return INT_MIN;
1582     s->lambda= pict->quality * 3/2;
1583     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1584     s->qlog+= delta_qlog;
1585     return delta_qlog;
1586 }
1587
1588 static void calculate_visual_weight(SnowContext *s, Plane *p){
1589     int width = p->width;
1590     int height= p->height;
1591     int level, orientation, x, y;
1592
1593     for(level=0; level<s->spatial_decomposition_count; level++){
1594         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1595             SubBand *b= &p->band[level][orientation];
1596             IDWTELEM *ibuf= b->ibuf;
1597             int64_t error=0;
1598
1599             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1600             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1601             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1602             for(y=0; y<height; y++){
1603                 for(x=0; x<width; x++){
1604                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1605                     error += d*d;
1606                 }
1607             }
1608
1609             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1610         }
1611     }
1612 }
1613
1614 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1615                         const AVFrame *pict, int *got_packet)
1616 {
1617     SnowContext *s = avctx->priv_data;
1618     RangeCoder * const c= &s->c;
1619     AVFrame *pic = &s->new_picture;
1620     const int width= s->avctx->width;
1621     const int height= s->avctx->height;
1622     int level, orientation, plane_index, i, y, ret;
1623     uint8_t rc_header_bak[sizeof(s->header_state)];
1624     uint8_t rc_block_bak[sizeof(s->block_state)];
1625
1626     if (!pkt->data &&
1627         (ret = av_new_packet(pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0) {
1628         av_log(avctx, AV_LOG_ERROR, "Error getting output packet.\n");
1629         return ret;
1630     }
1631
1632     ff_init_range_encoder(c, pkt->data, pkt->size);
1633     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1634
1635     for(i=0; i<3; i++){
1636         int shift= !!i;
1637         for(y=0; y<(height>>shift); y++)
1638             memcpy(&s->input_picture.data[i][y * s->input_picture.linesize[i]],
1639                    &pict->data[i][y * pict->linesize[i]],
1640                    width>>shift);
1641     }
1642     s->new_picture = *pict;
1643
1644     s->m.picture_number= avctx->frame_number;
1645     if(avctx->flags&CODEC_FLAG_PASS2){
1646         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1647         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1648         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1649             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1650             if (pic->quality < 0)
1651                 return -1;
1652         }
1653     }else{
1654         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1655         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1656     }
1657
1658     if(s->pass1_rc && avctx->frame_number == 0)
1659         pic->quality = 2*FF_QP2LAMBDA;
1660     if (pic->quality) {
1661         s->qlog   = qscale2qlog(pic->quality);
1662         s->lambda = pic->quality * 3/2;
1663     }
1664     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1665         s->qlog= LOSSLESS_QLOG;
1666         s->lambda = 0;
1667     }//else keep previous frame's qlog until after motion estimation
1668
1669     ff_snow_frame_start(s);
1670
1671     s->m.current_picture_ptr= &s->m.current_picture;
1672     s->m.last_picture.f.pts = s->m.current_picture.f.pts;
1673     s->m.current_picture.f.pts = pict->pts;
1674     if(pic->pict_type == AV_PICTURE_TYPE_P){
1675         int block_width = (width +15)>>4;
1676         int block_height= (height+15)>>4;
1677         int stride= s->current_picture.linesize[0];
1678
1679         assert(s->current_picture.data[0]);
1680         assert(s->last_picture[0].data[0]);
1681
1682         s->m.avctx= s->avctx;
1683         s->m.current_picture.f.data[0] = s->current_picture.data[0];
1684         s->m.   last_picture.f.data[0] = s->last_picture[0].data[0];
1685         s->m.    new_picture.f.data[0] = s->  input_picture.data[0];
1686         s->m.   last_picture_ptr= &s->m.   last_picture;
1687         s->m.linesize=
1688         s->m.   last_picture.f.linesize[0] =
1689         s->m.    new_picture.f.linesize[0] =
1690         s->m.current_picture.f.linesize[0] = stride;
1691         s->m.uvlinesize= s->current_picture.linesize[1];
1692         s->m.width = width;
1693         s->m.height= height;
1694         s->m.mb_width = block_width;
1695         s->m.mb_height= block_height;
1696         s->m.mb_stride=   s->m.mb_width+1;
1697         s->m.b8_stride= 2*s->m.mb_width+1;
1698         s->m.f_code=1;
1699         s->m.pict_type = pic->pict_type;
1700         s->m.me_method= s->avctx->me_method;
1701         s->m.me.scene_change_score=0;
1702         s->m.flags= s->avctx->flags;
1703         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1704         s->m.out_format= FMT_H263;
1705         s->m.unrestricted_mv= 1;
1706
1707         s->m.lambda = s->lambda;
1708         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1709         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1710
1711         s->m.dsp= s->dsp; //move
1712         ff_init_me(&s->m);
1713         s->dsp= s->m.dsp;
1714     }
1715
1716     if(s->pass1_rc){
1717         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1718         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1719     }
1720
1721 redo_frame:
1722
1723     if (pic->pict_type == AV_PICTURE_TYPE_I)
1724         s->spatial_decomposition_count= 5;
1725     else
1726         s->spatial_decomposition_count= 5;
1727
1728     s->m.pict_type = pic->pict_type;
1729     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1730
1731     ff_snow_common_init_after_header(avctx);
1732
1733     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1734         for(plane_index=0; plane_index<3; plane_index++){
1735             calculate_visual_weight(s, &s->plane[plane_index]);
1736         }
1737     }
1738
1739     encode_header(s);
1740     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1741     encode_blocks(s, 1);
1742     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1743
1744     for(plane_index=0; plane_index<3; plane_index++){
1745         Plane *p= &s->plane[plane_index];
1746         int w= p->width;
1747         int h= p->height;
1748         int x, y;
1749 //        int bits= put_bits_count(&s->c.pb);
1750
1751         if (!s->memc_only) {
1752             //FIXME optimize
1753             if(pict->data[plane_index]) //FIXME gray hack
1754                 for(y=0; y<h; y++){
1755                     for(x=0; x<w; x++){
1756                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1757                     }
1758                 }
1759             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1760
1761             if(   plane_index==0
1762                && pic->pict_type == AV_PICTURE_TYPE_P
1763                && !(avctx->flags&CODEC_FLAG_PASS2)
1764                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1765                 ff_init_range_encoder(c, pkt->data, pkt->size);
1766                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1767                 pic->pict_type= AV_PICTURE_TYPE_I;
1768                 s->keyframe=1;
1769                 s->current_picture.key_frame=1;
1770                 goto redo_frame;
1771             }
1772
1773             if(s->qlog == LOSSLESS_QLOG){
1774                 for(y=0; y<h; y++){
1775                     for(x=0; x<w; x++){
1776                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1777                     }
1778                 }
1779             }else{
1780                 for(y=0; y<h; y++){
1781                     for(x=0; x<w; x++){
1782                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1783                     }
1784                 }
1785             }
1786
1787             /*  if(QUANTIZE2)
1788                 dwt_quantize(s, p, s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type);
1789             else*/
1790                 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1791
1792             if(s->pass1_rc && plane_index==0){
1793                 int delta_qlog = ratecontrol_1pass(s, pic);
1794                 if (delta_qlog <= INT_MIN)
1795                     return -1;
1796                 if(delta_qlog){
1797                     //reordering qlog in the bitstream would eliminate this reset
1798                     ff_init_range_encoder(c, pkt->data, pkt->size);
1799                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1800                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1801                     encode_header(s);
1802                     encode_blocks(s, 0);
1803                 }
1804             }
1805
1806             for(level=0; level<s->spatial_decomposition_count; level++){
1807                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1808                     SubBand *b= &p->band[level][orientation];
1809
1810                     if(!QUANTIZE2)
1811                         quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1812                     if(orientation==0)
1813                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1814                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1815                     assert(b->parent==NULL || b->parent->stride == b->stride*2);
1816                     if(orientation==0)
1817                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1818                 }
1819             }
1820
1821             for(level=0; level<s->spatial_decomposition_count; level++){
1822                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1823                     SubBand *b= &p->band[level][orientation];
1824
1825                     dequantize(s, b, b->ibuf, b->stride);
1826                 }
1827             }
1828
1829             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1830             if(s->qlog == LOSSLESS_QLOG){
1831                 for(y=0; y<h; y++){
1832                     for(x=0; x<w; x++){
1833                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1834                     }
1835                 }
1836             }
1837             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1838         }else{
1839             //ME/MC only
1840             if(pic->pict_type == AV_PICTURE_TYPE_I){
1841                 for(y=0; y<h; y++){
1842                     for(x=0; x<w; x++){
1843                         s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
1844                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1845                     }
1846                 }
1847             }else{
1848                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1849                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1850             }
1851         }
1852         if(s->avctx->flags&CODEC_FLAG_PSNR){
1853             int64_t error= 0;
1854
1855             if(pict->data[plane_index]) //FIXME gray hack
1856                 for(y=0; y<h; y++){
1857                     for(x=0; x<w; x++){
1858                         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];
1859                         error += d*d;
1860                     }
1861                 }
1862             s->avctx->error[plane_index] += error;
1863             s->current_picture.error[plane_index] = error;
1864         }
1865
1866     }
1867
1868     update_last_header_values(s);
1869
1870     ff_snow_release_buffer(avctx);
1871
1872     s->current_picture.coded_picture_number = avctx->frame_number;
1873     s->current_picture.pict_type = pict->pict_type;
1874     s->current_picture.quality = pict->quality;
1875     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1876     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1877     s->m.current_picture.f.display_picture_number =
1878     s->m.current_picture.f.coded_picture_number   = avctx->frame_number;
1879     s->m.current_picture.f.quality                = pic->quality;
1880     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1881     if(s->pass1_rc)
1882         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1883             return -1;
1884     if(avctx->flags&CODEC_FLAG_PASS1)
1885         ff_write_pass1_stats(&s->m);
1886     s->m.last_pict_type = s->m.pict_type;
1887     avctx->frame_bits = s->m.frame_bits;
1888     avctx->mv_bits = s->m.mv_bits;
1889     avctx->misc_bits = s->m.misc_bits;
1890     avctx->p_tex_bits = s->m.p_tex_bits;
1891
1892     emms_c();
1893
1894     pkt->size = ff_rac_terminate(c);
1895     if (avctx->coded_frame->key_frame)
1896         pkt->flags |= AV_PKT_FLAG_KEY;
1897     *got_packet = 1;
1898
1899     return 0;
1900 }
1901
1902 static av_cold int encode_end(AVCodecContext *avctx)
1903 {
1904     SnowContext *s = avctx->priv_data;
1905
1906     ff_snow_common_end(s);
1907     if (s->input_picture.data[0])
1908         avctx->release_buffer(avctx, &s->input_picture);
1909     av_free(avctx->stats_out);
1910
1911     return 0;
1912 }
1913
1914 #define OFFSET(x) offsetof(SnowContext, x)
1915 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1916 static const AVOption options[] = {
1917     { "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 },
1918     { NULL },
1919 };
1920
1921 static const AVClass snowenc_class = {
1922     .class_name = "snow encoder",
1923     .item_name  = av_default_item_name,
1924     .option     = options,
1925     .version    = LIBAVUTIL_VERSION_INT,
1926 };
1927
1928 AVCodec ff_snow_encoder = {
1929     .name           = "snow",
1930     .type           = AVMEDIA_TYPE_VIDEO,
1931     .id             = AV_CODEC_ID_SNOW,
1932     .priv_data_size = sizeof(SnowContext),
1933     .init           = encode_init,
1934     .encode2        = encode_frame,
1935     .close          = encode_end,
1936     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1937     .priv_class     = &snowenc_class,
1938 };