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