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