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/libm.h"
23 #include "libavutil/log.h"
24 #include "libavutil/opt.h"
25 #include "libavutil/pixdesc.h"
26 #include "avcodec.h"
27 #include "internal.h"
28 #include "packet_internal.h"
29 #include "snow_dwt.h"
30 #include "snow.h"
31
32 #include "rangecoder.h"
33 #include "mathops.h"
34
35 #include "mpegvideo.h"
36 #include "h263.h"
37
encode_init(AVCodecContext * avctx)38 static av_cold int encode_init(AVCodecContext *avctx)
39 {
40 SnowContext *s = avctx->priv_data;
41 int plane_index, ret;
42 int i;
43
44 #if FF_API_PRIVATE_OPT
45 FF_DISABLE_DEPRECATION_WARNINGS
46 if (avctx->prediction_method)
47 s->pred = avctx->prediction_method;
48 FF_ENABLE_DEPRECATION_WARNINGS
49 #endif
50
51 if(s->pred == DWT_97
52 && (avctx->flags & AV_CODEC_FLAG_QSCALE)
53 && avctx->global_quality == 0){
54 av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
55 return AVERROR(EINVAL);
56 }
57
58 s->spatial_decomposition_type= s->pred; //FIXME add decorrelator type r transform_type
59
60 s->mv_scale = (avctx->flags & AV_CODEC_FLAG_QPEL) ? 2 : 4;
61 s->block_max_depth= (avctx->flags & AV_CODEC_FLAG_4MV ) ? 1 : 0;
62
63 for(plane_index=0; plane_index<3; plane_index++){
64 s->plane[plane_index].diag_mc= 1;
65 s->plane[plane_index].htaps= 6;
66 s->plane[plane_index].hcoeff[0]= 40;
67 s->plane[plane_index].hcoeff[1]= -10;
68 s->plane[plane_index].hcoeff[2]= 2;
69 s->plane[plane_index].fast_mc= 1;
70 }
71
72 if ((ret = ff_snow_common_init(avctx)) < 0) {
73 return ret;
74 }
75 ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
76
77 ff_snow_alloc_blocks(s);
78
79 s->version=0;
80
81 s->m.avctx = avctx;
82 s->m.bit_rate= avctx->bit_rate;
83 s->m.lmin = avctx->mb_lmin;
84 s->m.lmax = avctx->mb_lmax;
85 s->m.mb_num = (avctx->width * avctx->height + 255) / 256; // For ratecontrol
86
87 s->m.me.temp =
88 s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
89 s->m.me.map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
90 s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
91 s->m.sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
92 if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.sc.obmc_scratchpad)
93 return AVERROR(ENOMEM);
94
95 ff_h263_encode_init(&s->m); //mv_penalty
96
97 s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
98
99 if(avctx->flags&AV_CODEC_FLAG_PASS1){
100 if(!avctx->stats_out)
101 avctx->stats_out = av_mallocz(256);
102
103 if (!avctx->stats_out)
104 return AVERROR(ENOMEM);
105 }
106 if((avctx->flags&AV_CODEC_FLAG_PASS2) || !(avctx->flags&AV_CODEC_FLAG_QSCALE)){
107 ret = ff_rate_control_init(&s->m);
108 if(ret < 0)
109 return ret;
110 }
111 s->pass1_rc= !(avctx->flags & (AV_CODEC_FLAG_QSCALE|AV_CODEC_FLAG_PASS2));
112
113 switch(avctx->pix_fmt){
114 case AV_PIX_FMT_YUV444P:
115 // case AV_PIX_FMT_YUV422P:
116 case AV_PIX_FMT_YUV420P:
117 // case AV_PIX_FMT_YUV411P:
118 case AV_PIX_FMT_YUV410P:
119 s->nb_planes = 3;
120 s->colorspace_type= 0;
121 break;
122 case AV_PIX_FMT_GRAY8:
123 s->nb_planes = 1;
124 s->colorspace_type = 1;
125 break;
126 /* case AV_PIX_FMT_RGB32:
127 s->colorspace= 1;
128 break;*/
129 default:
130 av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
131 return AVERROR_PATCHWELCOME;
132 }
133
134 ret = av_pix_fmt_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift,
135 &s->chroma_v_shift);
136 if (ret) {
137 av_log(avctx, AV_LOG_ERROR, "pixel format invalid or unknown\n");
138 return ret;
139 }
140
141 ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
142 ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
143
144 s->input_picture = av_frame_alloc();
145 if (!s->input_picture)
146 return AVERROR(ENOMEM);
147
148 if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
149 return ret;
150
151 if(s->motion_est == FF_ME_ITER){
152 int size= s->b_width * s->b_height << 2*s->block_max_depth;
153 for(i=0; i<s->max_ref_frames; i++){
154 s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
155 s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
156 if (!s->ref_mvs[i] || !s->ref_scores[i])
157 return AVERROR(ENOMEM);
158 }
159 }
160
161 return 0;
162 }
163
164 //near copy & paste from dsputil, FIXME
pix_sum(uint8_t * pix,int line_size,int w,int h)165 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
166 {
167 int s, i, j;
168
169 s = 0;
170 for (i = 0; i < h; i++) {
171 for (j = 0; j < w; j++) {
172 s += pix[0];
173 pix ++;
174 }
175 pix += line_size - w;
176 }
177 return s;
178 }
179
180 //near copy & paste from dsputil, FIXME
pix_norm1(uint8_t * pix,int line_size,int w)181 static int pix_norm1(uint8_t * pix, int line_size, int w)
182 {
183 int s, i, j;
184 const uint32_t *sq = ff_square_tab + 256;
185
186 s = 0;
187 for (i = 0; i < w; i++) {
188 for (j = 0; j < w; j ++) {
189 s += sq[pix[0]];
190 pix ++;
191 }
192 pix += line_size - w;
193 }
194 return s;
195 }
196
get_penalty_factor(int lambda,int lambda2,int type)197 static inline int get_penalty_factor(int lambda, int lambda2, int type){
198 switch(type&0xFF){
199 default:
200 case FF_CMP_SAD:
201 return lambda>>FF_LAMBDA_SHIFT;
202 case FF_CMP_DCT:
203 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
204 case FF_CMP_W53:
205 return (4*lambda)>>(FF_LAMBDA_SHIFT);
206 case FF_CMP_W97:
207 return (2*lambda)>>(FF_LAMBDA_SHIFT);
208 case FF_CMP_SATD:
209 case FF_CMP_DCT264:
210 return (2*lambda)>>FF_LAMBDA_SHIFT;
211 case FF_CMP_RD:
212 case FF_CMP_PSNR:
213 case FF_CMP_SSE:
214 case FF_CMP_NSSE:
215 return lambda2>>FF_LAMBDA_SHIFT;
216 case FF_CMP_BIT:
217 return 1;
218 }
219 }
220
221 //FIXME copy&paste
222 #define P_LEFT P[1]
223 #define P_TOP P[2]
224 #define P_TOPRIGHT P[3]
225 #define P_MEDIAN P[4]
226 #define P_MV1 P[9]
227 #define FLAG_QPEL 1 //must be 1
228
encode_q_branch(SnowContext * s,int level,int x,int y)229 static int encode_q_branch(SnowContext *s, int level, int x, int y){
230 uint8_t p_buffer[1024];
231 uint8_t i_buffer[1024];
232 uint8_t p_state[sizeof(s->block_state)];
233 uint8_t i_state[sizeof(s->block_state)];
234 RangeCoder pc, ic;
235 uint8_t *pbbak= s->c.bytestream;
236 uint8_t *pbbak_start= s->c.bytestream_start;
237 int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
238 const int w= s->b_width << s->block_max_depth;
239 const int h= s->b_height << s->block_max_depth;
240 const int rem_depth= s->block_max_depth - level;
241 const int index= (x + y*w) << rem_depth;
242 const int block_w= 1<<(LOG2_MB_SIZE - level);
243 int trx= (x+1)<<rem_depth;
244 int try= (y+1)<<rem_depth;
245 const BlockNode *left = x ? &s->block[index-1] : &null_block;
246 const BlockNode *top = y ? &s->block[index-w] : &null_block;
247 const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
248 const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
249 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
250 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
251 int pl = left->color[0];
252 int pcb= left->color[1];
253 int pcr= left->color[2];
254 int pmx, pmy;
255 int mx=0, my=0;
256 int l,cr,cb;
257 const int stride= s->current_picture->linesize[0];
258 const int uvstride= s->current_picture->linesize[1];
259 uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y* stride)*block_w,
260 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
261 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
262 int P[10][2];
263 int16_t last_mv[3][2];
264 int qpel= !!(s->avctx->flags & AV_CODEC_FLAG_QPEL); //unused
265 const int shift= 1+qpel;
266 MotionEstContext *c= &s->m.me;
267 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
268 int mx_context= av_log2(2*FFABS(left->mx - top->mx));
269 int my_context= av_log2(2*FFABS(left->my - top->my));
270 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
271 int ref, best_ref, ref_score, ref_mx, ref_my;
272
273 av_assert0(sizeof(s->block_state) >= 256);
274 if(s->keyframe){
275 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
276 return 0;
277 }
278
279 // clip predictors / edge ?
280
281 P_LEFT[0]= left->mx;
282 P_LEFT[1]= left->my;
283 P_TOP [0]= top->mx;
284 P_TOP [1]= top->my;
285 P_TOPRIGHT[0]= tr->mx;
286 P_TOPRIGHT[1]= tr->my;
287
288 last_mv[0][0]= s->block[index].mx;
289 last_mv[0][1]= s->block[index].my;
290 last_mv[1][0]= right->mx;
291 last_mv[1][1]= right->my;
292 last_mv[2][0]= bottom->mx;
293 last_mv[2][1]= bottom->my;
294
295 s->m.mb_stride=2;
296 s->m.mb_x=
297 s->m.mb_y= 0;
298 c->skip= 0;
299
300 av_assert1(c-> stride == stride);
301 av_assert1(c->uvstride == uvstride);
302
303 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
304 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
305 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
306 c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_DMV;
307
308 c->xmin = - x*block_w - 16+3;
309 c->ymin = - y*block_w - 16+3;
310 c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
311 c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
312
313 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
314 if(P_LEFT[1] > (c->ymax<<shift)) P_LEFT[1] = (c->ymax<<shift);
315 if(P_TOP[0] > (c->xmax<<shift)) P_TOP[0] = (c->xmax<<shift);
316 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
317 if(P_TOPRIGHT[0] < (c->xmin * (1<<shift))) P_TOPRIGHT[0]= (c->xmin * (1<<shift));
318 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
319 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
320
321 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
322 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
323
324 if (!y) {
325 c->pred_x= P_LEFT[0];
326 c->pred_y= P_LEFT[1];
327 } else {
328 c->pred_x = P_MEDIAN[0];
329 c->pred_y = P_MEDIAN[1];
330 }
331
332 score= INT_MAX;
333 best_ref= 0;
334 for(ref=0; ref<s->ref_frames; ref++){
335 init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
336
337 ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
338 (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
339
340 av_assert2(ref_mx >= c->xmin);
341 av_assert2(ref_mx <= c->xmax);
342 av_assert2(ref_my >= c->ymin);
343 av_assert2(ref_my <= c->ymax);
344
345 ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
346 ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
347 ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
348 if(s->ref_mvs[ref]){
349 s->ref_mvs[ref][index][0]= ref_mx;
350 s->ref_mvs[ref][index][1]= ref_my;
351 s->ref_scores[ref][index]= ref_score;
352 }
353 if(score > ref_score){
354 score= ref_score;
355 best_ref= ref;
356 mx= ref_mx;
357 my= ref_my;
358 }
359 }
360 //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
361
362 // subpel search
363 base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
364 pc= s->c;
365 pc.bytestream_start=
366 pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
367 memcpy(p_state, s->block_state, sizeof(s->block_state));
368
369 if(level!=s->block_max_depth)
370 put_rac(&pc, &p_state[4 + s_context], 1);
371 put_rac(&pc, &p_state[1 + left->type + top->type], 0);
372 if(s->ref_frames > 1)
373 put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
374 pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
375 put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
376 put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
377 p_len= pc.bytestream - pc.bytestream_start;
378 score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
379
380 block_s= block_w*block_w;
381 sum = pix_sum(current_data[0], stride, block_w, block_w);
382 l= (sum + block_s/2)/block_s;
383 iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
384
385 if (s->nb_planes > 2) {
386 block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
387 sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
388 cb= (sum + block_s/2)/block_s;
389 // iscore += pix_norm1(¤t_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
390 sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
391 cr= (sum + block_s/2)/block_s;
392 // iscore += pix_norm1(¤t_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
393 }else
394 cb = cr = 0;
395
396 ic= s->c;
397 ic.bytestream_start=
398 ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
399 memcpy(i_state, s->block_state, sizeof(s->block_state));
400 if(level!=s->block_max_depth)
401 put_rac(&ic, &i_state[4 + s_context], 1);
402 put_rac(&ic, &i_state[1 + left->type + top->type], 1);
403 put_symbol(&ic, &i_state[32], l-pl , 1);
404 if (s->nb_planes > 2) {
405 put_symbol(&ic, &i_state[64], cb-pcb, 1);
406 put_symbol(&ic, &i_state[96], cr-pcr, 1);
407 }
408 i_len= ic.bytestream - ic.bytestream_start;
409 iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
410
411 av_assert1(iscore < 255*255*256 + s->lambda2*10);
412 av_assert1(iscore >= 0);
413 av_assert1(l>=0 && l<=255);
414 av_assert1(pl>=0 && pl<=255);
415
416 if(level==0){
417 int varc= iscore >> 8;
418 int vard= score >> 8;
419 if (vard <= 64 || vard < varc)
420 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
421 else
422 c->scene_change_score+= s->m.qscale;
423 }
424
425 if(level!=s->block_max_depth){
426 put_rac(&s->c, &s->block_state[4 + s_context], 0);
427 score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
428 score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
429 score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
430 score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
431 score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
432
433 if(score2 < score && score2 < iscore)
434 return score2;
435 }
436
437 if(iscore < score){
438 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
439 memcpy(pbbak, i_buffer, i_len);
440 s->c= ic;
441 s->c.bytestream_start= pbbak_start;
442 s->c.bytestream= pbbak + i_len;
443 set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
444 memcpy(s->block_state, i_state, sizeof(s->block_state));
445 return iscore;
446 }else{
447 memcpy(pbbak, p_buffer, p_len);
448 s->c= pc;
449 s->c.bytestream_start= pbbak_start;
450 s->c.bytestream= pbbak + p_len;
451 set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
452 memcpy(s->block_state, p_state, sizeof(s->block_state));
453 return score;
454 }
455 }
456
encode_q_branch2(SnowContext * s,int level,int x,int y)457 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
458 const int w= s->b_width << s->block_max_depth;
459 const int rem_depth= s->block_max_depth - level;
460 const int index= (x + y*w) << rem_depth;
461 int trx= (x+1)<<rem_depth;
462 BlockNode *b= &s->block[index];
463 const BlockNode *left = x ? &s->block[index-1] : &null_block;
464 const BlockNode *top = y ? &s->block[index-w] : &null_block;
465 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
466 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
467 int pl = left->color[0];
468 int pcb= left->color[1];
469 int pcr= left->color[2];
470 int pmx, pmy;
471 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
472 int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
473 int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
474 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
475
476 if(s->keyframe){
477 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
478 return;
479 }
480
481 if(level!=s->block_max_depth){
482 if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
483 put_rac(&s->c, &s->block_state[4 + s_context], 1);
484 }else{
485 put_rac(&s->c, &s->block_state[4 + s_context], 0);
486 encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
487 encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
488 encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
489 encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
490 return;
491 }
492 }
493 if(b->type & BLOCK_INTRA){
494 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
495 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
496 put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
497 if (s->nb_planes > 2) {
498 put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
499 put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
500 }
501 set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
502 }else{
503 pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
504 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
505 if(s->ref_frames > 1)
506 put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
507 put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
508 put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
509 set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
510 }
511 }
512
get_dc(SnowContext * s,int mb_x,int mb_y,int plane_index)513 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
514 int i, x2, y2;
515 Plane *p= &s->plane[plane_index];
516 const int block_size = MB_SIZE >> s->block_max_depth;
517 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
518 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
519 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
520 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
521 const int ref_stride= s->current_picture->linesize[plane_index];
522 uint8_t *src= s-> input_picture->data[plane_index];
523 IDWTELEM *dst= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
524 const int b_stride = s->b_width << s->block_max_depth;
525 const int w= p->width;
526 const int h= p->height;
527 int index= mb_x + mb_y*b_stride;
528 BlockNode *b= &s->block[index];
529 BlockNode backup= *b;
530 int ab=0;
531 int aa=0;
532
533 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
534
535 b->type|= BLOCK_INTRA;
536 b->color[plane_index]= 0;
537 memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
538
539 for(i=0; i<4; i++){
540 int mb_x2= mb_x + (i &1) - 1;
541 int mb_y2= mb_y + (i>>1) - 1;
542 int x= block_w*mb_x2 + block_w/2;
543 int y= block_h*mb_y2 + block_h/2;
544
545 add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
546 x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
547
548 for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
549 for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
550 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
551 int obmc_v= obmc[index];
552 int d;
553 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
554 if(x<0) obmc_v += obmc[index + block_w];
555 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
556 if(x+block_w>w) obmc_v += obmc[index - block_w];
557 //FIXME precalculate this or simplify it somehow else
558
559 d = -dst[index] + (1<<(FRAC_BITS-1));
560 dst[index] = d;
561 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
562 aa += obmc_v * obmc_v; //FIXME precalculate this
563 }
564 }
565 }
566 *b= backup;
567
568 return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
569 }
570
get_block_bits(SnowContext * s,int x,int y,int w)571 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
572 const int b_stride = s->b_width << s->block_max_depth;
573 const int b_height = s->b_height<< s->block_max_depth;
574 int index= x + y*b_stride;
575 const BlockNode *b = &s->block[index];
576 const BlockNode *left = x ? &s->block[index-1] : &null_block;
577 const BlockNode *top = y ? &s->block[index-b_stride] : &null_block;
578 const BlockNode *tl = y && x ? &s->block[index-b_stride-1] : left;
579 const BlockNode *tr = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
580 int dmx, dmy;
581 // int mx_context= av_log2(2*FFABS(left->mx - top->mx));
582 // int my_context= av_log2(2*FFABS(left->my - top->my));
583
584 if(x<0 || x>=b_stride || y>=b_height)
585 return 0;
586 /*
587 1 0 0
588 01X 1-2 1
589 001XX 3-6 2-3
590 0001XXX 7-14 4-7
591 00001XXXX 15-30 8-15
592 */
593 //FIXME try accurate rate
594 //FIXME intra and inter predictors if surrounding blocks are not the same type
595 if(b->type & BLOCK_INTRA){
596 return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
597 + av_log2(2*FFABS(left->color[1] - b->color[1]))
598 + av_log2(2*FFABS(left->color[2] - b->color[2])));
599 }else{
600 pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
601 dmx-= b->mx;
602 dmy-= b->my;
603 return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
604 + av_log2(2*FFABS(dmy))
605 + av_log2(2*b->ref));
606 }
607 }
608
get_block_rd(SnowContext * s,int mb_x,int mb_y,int plane_index,uint8_t (* obmc_edged)[MB_SIZE * 2])609 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
610 Plane *p= &s->plane[plane_index];
611 const int block_size = MB_SIZE >> s->block_max_depth;
612 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
613 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
614 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
615 const int ref_stride= s->current_picture->linesize[plane_index];
616 uint8_t *dst= s->current_picture->data[plane_index];
617 uint8_t *src= s-> input_picture->data[plane_index];
618 IDWTELEM *pred= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
619 uint8_t *cur = s->scratchbuf;
620 uint8_t *tmp = s->emu_edge_buffer;
621 const int b_stride = s->b_width << s->block_max_depth;
622 const int b_height = s->b_height<< s->block_max_depth;
623 const int w= p->width;
624 const int h= p->height;
625 int distortion;
626 int rate= 0;
627 const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
628 int sx= block_w*mb_x - block_w/2;
629 int sy= block_h*mb_y - block_h/2;
630 int x0= FFMAX(0,-sx);
631 int y0= FFMAX(0,-sy);
632 int x1= FFMIN(block_w*2, w-sx);
633 int y1= FFMIN(block_h*2, h-sy);
634 int i,x,y;
635
636 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
637
638 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);
639
640 for(y=y0; y<y1; y++){
641 const uint8_t *obmc1= obmc_edged[y];
642 const IDWTELEM *pred1 = pred + y*obmc_stride;
643 uint8_t *cur1 = cur + y*ref_stride;
644 uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
645 for(x=x0; x<x1; x++){
646 #if FRAC_BITS >= LOG2_OBMC_MAX
647 int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
648 #else
649 int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
650 #endif
651 v = (v + pred1[x]) >> FRAC_BITS;
652 if(v&(~255)) v= ~(v>>31);
653 dst1[x] = v;
654 }
655 }
656
657 /* copy the regions where obmc[] = (uint8_t)256 */
658 if(LOG2_OBMC_MAX == 8
659 && (mb_x == 0 || mb_x == b_stride-1)
660 && (mb_y == 0 || mb_y == b_height-1)){
661 if(mb_x == 0)
662 x1 = block_w;
663 else
664 x0 = block_w;
665 if(mb_y == 0)
666 y1 = block_h;
667 else
668 y0 = block_h;
669 for(y=y0; y<y1; y++)
670 memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
671 }
672
673 if(block_w==16){
674 /* FIXME rearrange dsputil to fit 32x32 cmp functions */
675 /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
676 /* FIXME cmps overlap but do not cover the wavelet's whole support.
677 * So improving the score of one block is not strictly guaranteed
678 * to improve the score of the whole frame, thus iterative motion
679 * estimation does not always converge. */
680 if(s->avctx->me_cmp == FF_CMP_W97)
681 distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
682 else if(s->avctx->me_cmp == FF_CMP_W53)
683 distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
684 else{
685 distortion = 0;
686 for(i=0; i<4; i++){
687 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
688 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
689 }
690 }
691 }else{
692 av_assert2(block_w==8);
693 distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
694 }
695
696 if(plane_index==0){
697 for(i=0; i<4; i++){
698 /* ..RRr
699 * .RXx.
700 * rxx..
701 */
702 rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
703 }
704 if(mb_x == b_stride-2)
705 rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
706 }
707 return distortion + rate*penalty_factor;
708 }
709
get_4block_rd(SnowContext * s,int mb_x,int mb_y,int plane_index)710 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
711 int i, y2;
712 Plane *p= &s->plane[plane_index];
713 const int block_size = MB_SIZE >> s->block_max_depth;
714 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
715 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
716 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
717 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
718 const int ref_stride= s->current_picture->linesize[plane_index];
719 uint8_t *dst= s->current_picture->data[plane_index];
720 uint8_t *src= s-> input_picture->data[plane_index];
721 //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
722 // const has only been removed from zero_dst to suppress a warning
723 static IDWTELEM zero_dst[4096]; //FIXME
724 const int b_stride = s->b_width << s->block_max_depth;
725 const int w= p->width;
726 const int h= p->height;
727 int distortion= 0;
728 int rate= 0;
729 const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
730
731 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
732
733 for(i=0; i<9; i++){
734 int mb_x2= mb_x + (i%3) - 1;
735 int mb_y2= mb_y + (i/3) - 1;
736 int x= block_w*mb_x2 + block_w/2;
737 int y= block_h*mb_y2 + block_h/2;
738
739 add_yblock(s, 0, NULL, zero_dst, dst, obmc,
740 x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
741
742 //FIXME find a cleaner/simpler way to skip the outside stuff
743 for(y2= y; y2<0; y2++)
744 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
745 for(y2= h; y2<y+block_h; y2++)
746 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
747 if(x<0){
748 for(y2= y; y2<y+block_h; y2++)
749 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
750 }
751 if(x+block_w > w){
752 for(y2= y; y2<y+block_h; y2++)
753 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
754 }
755
756 av_assert1(block_w== 8 || block_w==16);
757 distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
758 }
759
760 if(plane_index==0){
761 BlockNode *b= &s->block[mb_x+mb_y*b_stride];
762 int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
763
764 /* ..RRRr
765 * .RXXx.
766 * .RXXx.
767 * rxxx.
768 */
769 if(merged)
770 rate = get_block_bits(s, mb_x, mb_y, 2);
771 for(i=merged?4:0; i<9; i++){
772 static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
773 rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
774 }
775 }
776 return distortion + rate*penalty_factor;
777 }
778
encode_subband_c0run(SnowContext * s,SubBand * b,const IDWTELEM * src,const IDWTELEM * parent,int stride,int orientation)779 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
780 const int w= b->width;
781 const int h= b->height;
782 int x, y;
783
784 if(1){
785 int run=0;
786 int *runs = s->run_buffer;
787 int run_index=0;
788 int max_index;
789
790 for(y=0; y<h; y++){
791 for(x=0; x<w; x++){
792 int v, p=0;
793 int /*ll=0, */l=0, lt=0, t=0, rt=0;
794 v= src[x + y*stride];
795
796 if(y){
797 t= src[x + (y-1)*stride];
798 if(x){
799 lt= src[x - 1 + (y-1)*stride];
800 }
801 if(x + 1 < w){
802 rt= src[x + 1 + (y-1)*stride];
803 }
804 }
805 if(x){
806 l= src[x - 1 + y*stride];
807 /*if(x > 1){
808 if(orientation==1) ll= src[y + (x-2)*stride];
809 else ll= src[x - 2 + y*stride];
810 }*/
811 }
812 if(parent){
813 int px= x>>1;
814 int py= y>>1;
815 if(px<b->parent->width && py<b->parent->height)
816 p= parent[px + py*2*stride];
817 }
818 if(!(/*ll|*/l|lt|t|rt|p)){
819 if(v){
820 runs[run_index++]= run;
821 run=0;
822 }else{
823 run++;
824 }
825 }
826 }
827 }
828 max_index= run_index;
829 runs[run_index++]= run;
830 run_index=0;
831 run= runs[run_index++];
832
833 put_symbol2(&s->c, b->state[30], max_index, 0);
834 if(run_index <= max_index)
835 put_symbol2(&s->c, b->state[1], run, 3);
836
837 for(y=0; y<h; y++){
838 if(s->c.bytestream_end - s->c.bytestream < w*40){
839 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
840 return AVERROR(ENOMEM);
841 }
842 for(x=0; x<w; x++){
843 int v, p=0;
844 int /*ll=0, */l=0, lt=0, t=0, rt=0;
845 v= src[x + y*stride];
846
847 if(y){
848 t= src[x + (y-1)*stride];
849 if(x){
850 lt= src[x - 1 + (y-1)*stride];
851 }
852 if(x + 1 < w){
853 rt= src[x + 1 + (y-1)*stride];
854 }
855 }
856 if(x){
857 l= src[x - 1 + y*stride];
858 /*if(x > 1){
859 if(orientation==1) ll= src[y + (x-2)*stride];
860 else ll= src[x - 2 + y*stride];
861 }*/
862 }
863 if(parent){
864 int px= x>>1;
865 int py= y>>1;
866 if(px<b->parent->width && py<b->parent->height)
867 p= parent[px + py*2*stride];
868 }
869 if(/*ll|*/l|lt|t|rt|p){
870 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
871
872 put_rac(&s->c, &b->state[0][context], !!v);
873 }else{
874 if(!run){
875 run= runs[run_index++];
876
877 if(run_index <= max_index)
878 put_symbol2(&s->c, b->state[1], run, 3);
879 av_assert2(v);
880 }else{
881 run--;
882 av_assert2(!v);
883 }
884 }
885 if(v){
886 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
887 int l2= 2*FFABS(l) + (l<0);
888 int t2= 2*FFABS(t) + (t<0);
889
890 put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
891 put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
892 }
893 }
894 }
895 }
896 return 0;
897 }
898
encode_subband(SnowContext * s,SubBand * b,const IDWTELEM * src,const IDWTELEM * parent,int stride,int orientation)899 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
900 // encode_subband_qtree(s, b, src, parent, stride, orientation);
901 // encode_subband_z0run(s, b, src, parent, stride, orientation);
902 return encode_subband_c0run(s, b, src, parent, stride, orientation);
903 // encode_subband_dzr(s, b, src, parent, stride, orientation);
904 }
905
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)906 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){
907 const int b_stride= s->b_width << s->block_max_depth;
908 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
909 BlockNode backup= *block;
910 unsigned value;
911 int rd, index;
912
913 av_assert2(mb_x>=0 && mb_y>=0);
914 av_assert2(mb_x<b_stride);
915
916 if(intra){
917 block->color[0] = p[0];
918 block->color[1] = p[1];
919 block->color[2] = p[2];
920 block->type |= BLOCK_INTRA;
921 }else{
922 index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
923 value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
924 if(s->me_cache[index] == value)
925 return 0;
926 s->me_cache[index]= value;
927
928 block->mx= p[0];
929 block->my= p[1];
930 block->type &= ~BLOCK_INTRA;
931 }
932
933 rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged) + s->intra_penalty * !!intra;
934
935 //FIXME chroma
936 if(rd < *best_rd){
937 *best_rd= rd;
938 return 1;
939 }else{
940 *block= backup;
941 return 0;
942 }
943 }
944
945 /* special case for int[2] args we discard afterwards,
946 * fixes compilation problem with gcc 2.95 */
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)947 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){
948 int p[2] = {p0, p1};
949 return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
950 }
951
check_4block_inter(SnowContext * s,int mb_x,int mb_y,int p0,int p1,int ref,int * best_rd)952 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){
953 const int b_stride= s->b_width << s->block_max_depth;
954 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
955 BlockNode backup[4];
956 unsigned value;
957 int rd, index;
958
959 /* We don't initialize backup[] during variable declaration, because
960 * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
961 * 'int16_t'". */
962 backup[0] = block[0];
963 backup[1] = block[1];
964 backup[2] = block[b_stride];
965 backup[3] = block[b_stride + 1];
966
967 av_assert2(mb_x>=0 && mb_y>=0);
968 av_assert2(mb_x<b_stride);
969 av_assert2(((mb_x|mb_y)&1) == 0);
970
971 index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
972 value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
973 if(s->me_cache[index] == value)
974 return 0;
975 s->me_cache[index]= value;
976
977 block->mx= p0;
978 block->my= p1;
979 block->ref= ref;
980 block->type &= ~BLOCK_INTRA;
981 block[1]= block[b_stride]= block[b_stride+1]= *block;
982
983 rd= get_4block_rd(s, mb_x, mb_y, 0);
984
985 //FIXME chroma
986 if(rd < *best_rd){
987 *best_rd= rd;
988 return 1;
989 }else{
990 block[0]= backup[0];
991 block[1]= backup[1];
992 block[b_stride]= backup[2];
993 block[b_stride+1]= backup[3];
994 return 0;
995 }
996 }
997
iterative_me(SnowContext * s)998 static void iterative_me(SnowContext *s){
999 int pass, mb_x, mb_y;
1000 const int b_width = s->b_width << s->block_max_depth;
1001 const int b_height= s->b_height << s->block_max_depth;
1002 const int b_stride= b_width;
1003 int color[3];
1004
1005 {
1006 RangeCoder r = s->c;
1007 uint8_t state[sizeof(s->block_state)];
1008 memcpy(state, s->block_state, sizeof(s->block_state));
1009 for(mb_y= 0; mb_y<s->b_height; mb_y++)
1010 for(mb_x= 0; mb_x<s->b_width; mb_x++)
1011 encode_q_branch(s, 0, mb_x, mb_y);
1012 s->c = r;
1013 memcpy(s->block_state, state, sizeof(s->block_state));
1014 }
1015
1016 for(pass=0; pass<25; pass++){
1017 int change= 0;
1018
1019 for(mb_y= 0; mb_y<b_height; mb_y++){
1020 for(mb_x= 0; mb_x<b_width; mb_x++){
1021 int dia_change, i, j, ref;
1022 int best_rd= INT_MAX, ref_rd;
1023 BlockNode backup, ref_b;
1024 const int index= mb_x + mb_y * b_stride;
1025 BlockNode *block= &s->block[index];
1026 BlockNode *tb = mb_y ? &s->block[index-b_stride ] : NULL;
1027 BlockNode *lb = mb_x ? &s->block[index -1] : NULL;
1028 BlockNode *rb = mb_x+1<b_width ? &s->block[index +1] : NULL;
1029 BlockNode *bb = mb_y+1<b_height ? &s->block[index+b_stride ] : NULL;
1030 BlockNode *tlb= mb_x && mb_y ? &s->block[index-b_stride-1] : NULL;
1031 BlockNode *trb= mb_x+1<b_width && mb_y ? &s->block[index-b_stride+1] : NULL;
1032 BlockNode *blb= mb_x && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1033 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1034 const int b_w= (MB_SIZE >> s->block_max_depth);
1035 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1036
1037 if(pass && (block->type & BLOCK_OPT))
1038 continue;
1039 block->type |= BLOCK_OPT;
1040
1041 backup= *block;
1042
1043 if(!s->me_cache_generation)
1044 memset(s->me_cache, 0, sizeof(s->me_cache));
1045 s->me_cache_generation += 1<<22;
1046
1047 //FIXME precalculate
1048 {
1049 int x, y;
1050 for (y = 0; y < b_w * 2; y++)
1051 memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1052 if(mb_x==0)
1053 for(y=0; y<b_w*2; y++)
1054 memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1055 if(mb_x==b_stride-1)
1056 for(y=0; y<b_w*2; y++)
1057 memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1058 if(mb_y==0){
1059 for(x=0; x<b_w*2; x++)
1060 obmc_edged[0][x] += obmc_edged[b_w-1][x];
1061 for(y=1; y<b_w; y++)
1062 memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1063 }
1064 if(mb_y==b_height-1){
1065 for(x=0; x<b_w*2; x++)
1066 obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1067 for(y=b_w; y<b_w*2-1; y++)
1068 memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1069 }
1070 }
1071
1072 //skip stuff outside the picture
1073 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1074 uint8_t *src= s-> input_picture->data[0];
1075 uint8_t *dst= s->current_picture->data[0];
1076 const int stride= s->current_picture->linesize[0];
1077 const int block_w= MB_SIZE >> s->block_max_depth;
1078 const int block_h= MB_SIZE >> s->block_max_depth;
1079 const int sx= block_w*mb_x - block_w/2;
1080 const int sy= block_h*mb_y - block_h/2;
1081 const int w= s->plane[0].width;
1082 const int h= s->plane[0].height;
1083 int y;
1084
1085 for(y=sy; y<0; y++)
1086 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1087 for(y=h; y<sy+block_h*2; y++)
1088 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1089 if(sx<0){
1090 for(y=sy; y<sy+block_h*2; y++)
1091 memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1092 }
1093 if(sx+block_w*2 > w){
1094 for(y=sy; y<sy+block_h*2; y++)
1095 memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1096 }
1097 }
1098
1099 // intra(black) = neighbors' contribution to the current block
1100 for(i=0; i < s->nb_planes; i++)
1101 color[i]= get_dc(s, mb_x, mb_y, i);
1102
1103 // get previous score (cannot be cached due to OBMC)
1104 if(pass > 0 && (block->type&BLOCK_INTRA)){
1105 int color0[3]= {block->color[0], block->color[1], block->color[2]};
1106 check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1107 }else
1108 check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1109
1110 ref_b= *block;
1111 ref_rd= best_rd;
1112 for(ref=0; ref < s->ref_frames; ref++){
1113 int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1114 if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1115 continue;
1116 block->ref= ref;
1117 best_rd= INT_MAX;
1118
1119 check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1120 check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1121 if(tb)
1122 check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1123 if(lb)
1124 check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1125 if(rb)
1126 check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1127 if(bb)
1128 check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1129
1130 /* fullpel ME */
1131 //FIXME avoid subpel interpolation / round to nearest integer
1132 do{
1133 int newx = block->mx;
1134 int newy = block->my;
1135 int dia_size = s->iterative_dia_size ? s->iterative_dia_size : FFMAX(s->avctx->dia_size, 1);
1136 dia_change=0;
1137 for(i=0; i < dia_size; i++){
1138 for(j=0; j<i; j++){
1139 dia_change |= check_block_inter(s, mb_x, mb_y, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
1140 dia_change |= check_block_inter(s, mb_x, mb_y, newx-4*(i-j), newy-(4*j), obmc_edged, &best_rd);
1141 dia_change |= check_block_inter(s, mb_x, mb_y, newx-(4*j), newy+4*(i-j), obmc_edged, &best_rd);
1142 dia_change |= check_block_inter(s, mb_x, mb_y, newx+(4*j), newy-4*(i-j), obmc_edged, &best_rd);
1143 }
1144 }
1145 }while(dia_change);
1146 /* subpel ME */
1147 do{
1148 static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1149 dia_change=0;
1150 for(i=0; i<8; i++)
1151 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1152 }while(dia_change);
1153 //FIXME or try the standard 2 pass qpel or similar
1154
1155 mvr[0][0]= block->mx;
1156 mvr[0][1]= block->my;
1157 if(ref_rd > best_rd){
1158 ref_rd= best_rd;
1159 ref_b= *block;
1160 }
1161 }
1162 best_rd= ref_rd;
1163 *block= ref_b;
1164 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1165 //FIXME RD style color selection
1166 if(!same_block(block, &backup)){
1167 if(tb ) tb ->type &= ~BLOCK_OPT;
1168 if(lb ) lb ->type &= ~BLOCK_OPT;
1169 if(rb ) rb ->type &= ~BLOCK_OPT;
1170 if(bb ) bb ->type &= ~BLOCK_OPT;
1171 if(tlb) tlb->type &= ~BLOCK_OPT;
1172 if(trb) trb->type &= ~BLOCK_OPT;
1173 if(blb) blb->type &= ~BLOCK_OPT;
1174 if(brb) brb->type &= ~BLOCK_OPT;
1175 change ++;
1176 }
1177 }
1178 }
1179 av_log(s->avctx, AV_LOG_DEBUG, "pass:%d changed:%d\n", pass, change);
1180 if(!change)
1181 break;
1182 }
1183
1184 if(s->block_max_depth == 1){
1185 int change= 0;
1186 for(mb_y= 0; mb_y<b_height; mb_y+=2){
1187 for(mb_x= 0; mb_x<b_width; mb_x+=2){
1188 int i;
1189 int best_rd, init_rd;
1190 const int index= mb_x + mb_y * b_stride;
1191 BlockNode *b[4];
1192
1193 b[0]= &s->block[index];
1194 b[1]= b[0]+1;
1195 b[2]= b[0]+b_stride;
1196 b[3]= b[2]+1;
1197 if(same_block(b[0], b[1]) &&
1198 same_block(b[0], b[2]) &&
1199 same_block(b[0], b[3]))
1200 continue;
1201
1202 if(!s->me_cache_generation)
1203 memset(s->me_cache, 0, sizeof(s->me_cache));
1204 s->me_cache_generation += 1<<22;
1205
1206 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1207
1208 //FIXME more multiref search?
1209 check_4block_inter(s, mb_x, mb_y,
1210 (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1211 (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1212
1213 for(i=0; i<4; i++)
1214 if(!(b[i]->type&BLOCK_INTRA))
1215 check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1216
1217 if(init_rd != best_rd)
1218 change++;
1219 }
1220 }
1221 av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1222 }
1223 }
1224
encode_blocks(SnowContext * s,int search)1225 static void encode_blocks(SnowContext *s, int search){
1226 int x, y;
1227 int w= s->b_width;
1228 int h= s->b_height;
1229
1230 if(s->motion_est == FF_ME_ITER && !s->keyframe && search)
1231 iterative_me(s);
1232
1233 for(y=0; y<h; y++){
1234 if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1235 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1236 return;
1237 }
1238 for(x=0; x<w; x++){
1239 if(s->motion_est == FF_ME_ITER || !search)
1240 encode_q_branch2(s, 0, x, y);
1241 else
1242 encode_q_branch (s, 0, x, y);
1243 }
1244 }
1245 }
1246
quantize(SnowContext * s,SubBand * b,IDWTELEM * dst,DWTELEM * src,int stride,int bias)1247 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1248 const int w= b->width;
1249 const int h= b->height;
1250 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1251 const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1252 int x,y, thres1, thres2;
1253
1254 if(s->qlog == LOSSLESS_QLOG){
1255 for(y=0; y<h; y++)
1256 for(x=0; x<w; x++)
1257 dst[x + y*stride]= src[x + y*stride];
1258 return;
1259 }
1260
1261 bias= bias ? 0 : (3*qmul)>>3;
1262 thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1263 thres2= 2*thres1;
1264
1265 if(!bias){
1266 for(y=0; y<h; y++){
1267 for(x=0; x<w; x++){
1268 int i= src[x + y*stride];
1269
1270 if((unsigned)(i+thres1) > thres2){
1271 if(i>=0){
1272 i<<= QEXPSHIFT;
1273 i/= qmul; //FIXME optimize
1274 dst[x + y*stride]= i;
1275 }else{
1276 i= -i;
1277 i<<= QEXPSHIFT;
1278 i/= qmul; //FIXME optimize
1279 dst[x + y*stride]= -i;
1280 }
1281 }else
1282 dst[x + y*stride]= 0;
1283 }
1284 }
1285 }else{
1286 for(y=0; y<h; y++){
1287 for(x=0; x<w; x++){
1288 int i= src[x + y*stride];
1289
1290 if((unsigned)(i+thres1) > thres2){
1291 if(i>=0){
1292 i<<= QEXPSHIFT;
1293 i= (i + bias) / qmul; //FIXME optimize
1294 dst[x + y*stride]= i;
1295 }else{
1296 i= -i;
1297 i<<= QEXPSHIFT;
1298 i= (i + bias) / qmul; //FIXME optimize
1299 dst[x + y*stride]= -i;
1300 }
1301 }else
1302 dst[x + y*stride]= 0;
1303 }
1304 }
1305 }
1306 }
1307
dequantize(SnowContext * s,SubBand * b,IDWTELEM * src,int stride)1308 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1309 const int w= b->width;
1310 const int h= b->height;
1311 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1312 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1313 const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1314 int x,y;
1315
1316 if(s->qlog == LOSSLESS_QLOG) return;
1317
1318 for(y=0; y<h; y++){
1319 for(x=0; x<w; x++){
1320 int i= src[x + y*stride];
1321 if(i<0){
1322 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1323 }else if(i>0){
1324 src[x + y*stride]= (( i*qmul + qadd)>>(QEXPSHIFT));
1325 }
1326 }
1327 }
1328 }
1329
decorrelate(SnowContext * s,SubBand * b,IDWTELEM * src,int stride,int inverse,int use_median)1330 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1331 const int w= b->width;
1332 const int h= b->height;
1333 int x,y;
1334
1335 for(y=h-1; y>=0; y--){
1336 for(x=w-1; x>=0; x--){
1337 int i= x + y*stride;
1338
1339 if(x){
1340 if(use_median){
1341 if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1342 else src[i] -= src[i - 1];
1343 }else{
1344 if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1345 else src[i] -= src[i - 1];
1346 }
1347 }else{
1348 if(y) src[i] -= src[i - stride];
1349 }
1350 }
1351 }
1352 }
1353
correlate(SnowContext * s,SubBand * b,IDWTELEM * src,int stride,int inverse,int use_median)1354 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1355 const int w= b->width;
1356 const int h= b->height;
1357 int x,y;
1358
1359 for(y=0; y<h; y++){
1360 for(x=0; x<w; x++){
1361 int i= x + y*stride;
1362
1363 if(x){
1364 if(use_median){
1365 if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1366 else src[i] += src[i - 1];
1367 }else{
1368 if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1369 else src[i] += src[i - 1];
1370 }
1371 }else{
1372 if(y) src[i] += src[i - stride];
1373 }
1374 }
1375 }
1376 }
1377
encode_qlogs(SnowContext * s)1378 static void encode_qlogs(SnowContext *s){
1379 int plane_index, level, orientation;
1380
1381 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1382 for(level=0; level<s->spatial_decomposition_count; level++){
1383 for(orientation=level ? 1:0; orientation<4; orientation++){
1384 if(orientation==2) continue;
1385 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1386 }
1387 }
1388 }
1389 }
1390
encode_header(SnowContext * s)1391 static void encode_header(SnowContext *s){
1392 int plane_index, i;
1393 uint8_t kstate[32];
1394
1395 memset(kstate, MID_STATE, sizeof(kstate));
1396
1397 put_rac(&s->c, kstate, s->keyframe);
1398 if(s->keyframe || s->always_reset){
1399 ff_snow_reset_contexts(s);
1400 s->last_spatial_decomposition_type=
1401 s->last_qlog=
1402 s->last_qbias=
1403 s->last_mv_scale=
1404 s->last_block_max_depth= 0;
1405 for(plane_index=0; plane_index<2; plane_index++){
1406 Plane *p= &s->plane[plane_index];
1407 p->last_htaps=0;
1408 p->last_diag_mc=0;
1409 memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1410 }
1411 }
1412 if(s->keyframe){
1413 put_symbol(&s->c, s->header_state, s->version, 0);
1414 put_rac(&s->c, s->header_state, s->always_reset);
1415 put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1416 put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1417 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1418 put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1419 if (s->nb_planes > 2) {
1420 put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1421 put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1422 }
1423 put_rac(&s->c, s->header_state, s->spatial_scalability);
1424 // put_rac(&s->c, s->header_state, s->rate_scalability);
1425 put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1426
1427 encode_qlogs(s);
1428 }
1429
1430 if(!s->keyframe){
1431 int update_mc=0;
1432 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1433 Plane *p= &s->plane[plane_index];
1434 update_mc |= p->last_htaps != p->htaps;
1435 update_mc |= p->last_diag_mc != p->diag_mc;
1436 update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1437 }
1438 put_rac(&s->c, s->header_state, update_mc);
1439 if(update_mc){
1440 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1441 Plane *p= &s->plane[plane_index];
1442 put_rac(&s->c, s->header_state, p->diag_mc);
1443 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1444 for(i= p->htaps/2; i; i--)
1445 put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1446 }
1447 }
1448 if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1449 put_rac(&s->c, s->header_state, 1);
1450 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1451 encode_qlogs(s);
1452 }else
1453 put_rac(&s->c, s->header_state, 0);
1454 }
1455
1456 put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1457 put_symbol(&s->c, s->header_state, s->qlog - s->last_qlog , 1);
1458 put_symbol(&s->c, s->header_state, s->mv_scale - s->last_mv_scale, 1);
1459 put_symbol(&s->c, s->header_state, s->qbias - s->last_qbias , 1);
1460 put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1461
1462 }
1463
update_last_header_values(SnowContext * s)1464 static void update_last_header_values(SnowContext *s){
1465 int plane_index;
1466
1467 if(!s->keyframe){
1468 for(plane_index=0; plane_index<2; plane_index++){
1469 Plane *p= &s->plane[plane_index];
1470 p->last_diag_mc= p->diag_mc;
1471 p->last_htaps = p->htaps;
1472 memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1473 }
1474 }
1475
1476 s->last_spatial_decomposition_type = s->spatial_decomposition_type;
1477 s->last_qlog = s->qlog;
1478 s->last_qbias = s->qbias;
1479 s->last_mv_scale = s->mv_scale;
1480 s->last_block_max_depth = s->block_max_depth;
1481 s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1482 }
1483
qscale2qlog(int qscale)1484 static int qscale2qlog(int qscale){
1485 return lrint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1486 + 61*QROOT/8; ///< 64 > 60
1487 }
1488
ratecontrol_1pass(SnowContext * s,AVFrame * pict)1489 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1490 {
1491 /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1492 * FIXME we know exact mv bits at this point,
1493 * but ratecontrol isn't set up to include them. */
1494 uint32_t coef_sum= 0;
1495 int level, orientation, delta_qlog;
1496
1497 for(level=0; level<s->spatial_decomposition_count; level++){
1498 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1499 SubBand *b= &s->plane[0].band[level][orientation];
1500 IDWTELEM *buf= b->ibuf;
1501 const int w= b->width;
1502 const int h= b->height;
1503 const int stride= b->stride;
1504 const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1505 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1506 const int qdiv= (1<<16)/qmul;
1507 int x, y;
1508 //FIXME this is ugly
1509 for(y=0; y<h; y++)
1510 for(x=0; x<w; x++)
1511 buf[x+y*stride]= b->buf[x+y*stride];
1512 if(orientation==0)
1513 decorrelate(s, b, buf, stride, 1, 0);
1514 for(y=0; y<h; y++)
1515 for(x=0; x<w; x++)
1516 coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1517 }
1518 }
1519
1520 /* ugly, ratecontrol just takes a sqrt again */
1521 av_assert0(coef_sum < INT_MAX);
1522 coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1523
1524 if(pict->pict_type == AV_PICTURE_TYPE_I){
1525 s->m.current_picture.mb_var_sum= coef_sum;
1526 s->m.current_picture.mc_mb_var_sum= 0;
1527 }else{
1528 s->m.current_picture.mc_mb_var_sum= coef_sum;
1529 s->m.current_picture.mb_var_sum= 0;
1530 }
1531
1532 pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1533 if (pict->quality < 0)
1534 return INT_MIN;
1535 s->lambda= pict->quality * 3/2;
1536 delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1537 s->qlog+= delta_qlog;
1538 return delta_qlog;
1539 }
1540
calculate_visual_weight(SnowContext * s,Plane * p)1541 static void calculate_visual_weight(SnowContext *s, Plane *p){
1542 int width = p->width;
1543 int height= p->height;
1544 int level, orientation, x, y;
1545
1546 for(level=0; level<s->spatial_decomposition_count; level++){
1547 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1548 SubBand *b= &p->band[level][orientation];
1549 IDWTELEM *ibuf= b->ibuf;
1550 int64_t error=0;
1551
1552 memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1553 ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1554 ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1555 for(y=0; y<height; y++){
1556 for(x=0; x<width; x++){
1557 int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1558 error += d*d;
1559 }
1560 }
1561
1562 b->qlog= (int)(QROOT * log2(352256.0/sqrt(error)) + 0.5);
1563 }
1564 }
1565 }
1566
encode_frame(AVCodecContext * avctx,AVPacket * pkt,const AVFrame * pict,int * got_packet)1567 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1568 const AVFrame *pict, int *got_packet)
1569 {
1570 SnowContext *s = avctx->priv_data;
1571 RangeCoder * const c= &s->c;
1572 AVFrame *pic;
1573 const int width= s->avctx->width;
1574 const int height= s->avctx->height;
1575 int level, orientation, plane_index, i, y, ret;
1576 uint8_t rc_header_bak[sizeof(s->header_state)];
1577 uint8_t rc_block_bak[sizeof(s->block_state)];
1578
1579 if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + AV_INPUT_BUFFER_MIN_SIZE, 0)) < 0)
1580 return ret;
1581
1582 ff_init_range_encoder(c, pkt->data, pkt->size);
1583 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1584
1585 for(i=0; i < s->nb_planes; i++){
1586 int hshift= i ? s->chroma_h_shift : 0;
1587 int vshift= i ? s->chroma_v_shift : 0;
1588 for(y=0; y<AV_CEIL_RSHIFT(height, vshift); y++)
1589 memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1590 &pict->data[i][y * pict->linesize[i]],
1591 AV_CEIL_RSHIFT(width, hshift));
1592 s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1593 AV_CEIL_RSHIFT(width, hshift), AV_CEIL_RSHIFT(height, vshift),
1594 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1595 EDGE_TOP | EDGE_BOTTOM);
1596
1597 }
1598 emms_c();
1599 pic = s->input_picture;
1600 pic->pict_type = pict->pict_type;
1601 pic->quality = pict->quality;
1602
1603 s->m.picture_number= avctx->frame_number;
1604 if(avctx->flags&AV_CODEC_FLAG_PASS2){
1605 s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1606 s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1607 if(!(avctx->flags&AV_CODEC_FLAG_QSCALE)) {
1608 pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1609 if (pic->quality < 0)
1610 return -1;
1611 }
1612 }else{
1613 s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1614 s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1615 }
1616
1617 if(s->pass1_rc && avctx->frame_number == 0)
1618 pic->quality = 2*FF_QP2LAMBDA;
1619 if (pic->quality) {
1620 s->qlog = qscale2qlog(pic->quality);
1621 s->lambda = pic->quality * 3/2;
1622 }
1623 if (s->qlog < 0 || (!pic->quality && (avctx->flags & AV_CODEC_FLAG_QSCALE))) {
1624 s->qlog= LOSSLESS_QLOG;
1625 s->lambda = 0;
1626 }//else keep previous frame's qlog until after motion estimation
1627
1628 #if FF_API_CODED_FRAME
1629 FF_DISABLE_DEPRECATION_WARNINGS
1630 av_frame_unref(avctx->coded_frame);
1631 FF_ENABLE_DEPRECATION_WARNINGS
1632 #endif
1633
1634 if (s->current_picture->data[0]) {
1635 int w = s->avctx->width;
1636 int h = s->avctx->height;
1637
1638 #if FF_API_CODED_FRAME
1639 ret = av_frame_make_writable(s->current_picture);
1640 if (ret < 0)
1641 return ret;
1642 #endif
1643
1644 s->mpvencdsp.draw_edges(s->current_picture->data[0],
1645 s->current_picture->linesize[0], w , h ,
1646 EDGE_WIDTH , EDGE_WIDTH , EDGE_TOP | EDGE_BOTTOM);
1647 if (s->current_picture->data[2]) {
1648 s->mpvencdsp.draw_edges(s->current_picture->data[1],
1649 s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1650 EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1651 s->mpvencdsp.draw_edges(s->current_picture->data[2],
1652 s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1653 EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1654 }
1655 emms_c();
1656 }
1657
1658 ff_snow_frame_start(s);
1659 #if FF_API_CODED_FRAME
1660 FF_DISABLE_DEPRECATION_WARNINGS
1661 ret = av_frame_ref(avctx->coded_frame, s->current_picture);
1662 FF_ENABLE_DEPRECATION_WARNINGS
1663 #endif
1664 if (ret < 0)
1665 return ret;
1666
1667 s->m.current_picture_ptr= &s->m.current_picture;
1668 s->m.current_picture.f = s->current_picture;
1669 s->m.current_picture.f->pts = pict->pts;
1670 if(pic->pict_type == AV_PICTURE_TYPE_P){
1671 int block_width = (width +15)>>4;
1672 int block_height= (height+15)>>4;
1673 int stride= s->current_picture->linesize[0];
1674
1675 av_assert0(s->current_picture->data[0]);
1676 av_assert0(s->last_picture[0]->data[0]);
1677
1678 s->m.avctx= s->avctx;
1679 s->m. last_picture.f = s->last_picture[0];
1680 s->m. new_picture.f = s->input_picture;
1681 s->m. last_picture_ptr= &s->m. last_picture;
1682 s->m.linesize = stride;
1683 s->m.uvlinesize= s->current_picture->linesize[1];
1684 s->m.width = width;
1685 s->m.height= height;
1686 s->m.mb_width = block_width;
1687 s->m.mb_height= block_height;
1688 s->m.mb_stride= s->m.mb_width+1;
1689 s->m.b8_stride= 2*s->m.mb_width+1;
1690 s->m.f_code=1;
1691 s->m.pict_type = pic->pict_type;
1692 s->m.motion_est= s->motion_est;
1693 s->m.me.scene_change_score=0;
1694 s->m.me.dia_size = avctx->dia_size;
1695 s->m.quarter_sample= (s->avctx->flags & AV_CODEC_FLAG_QPEL)!=0;
1696 s->m.out_format= FMT_H263;
1697 s->m.unrestricted_mv= 1;
1698
1699 s->m.lambda = s->lambda;
1700 s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1701 s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1702
1703 s->m.mecc= s->mecc; //move
1704 s->m.qdsp= s->qdsp; //move
1705 s->m.hdsp = s->hdsp;
1706 ff_init_me(&s->m);
1707 s->hdsp = s->m.hdsp;
1708 s->mecc= s->m.mecc;
1709 }
1710
1711 if(s->pass1_rc){
1712 memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1713 memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1714 }
1715
1716 redo_frame:
1717
1718 s->spatial_decomposition_count= 5;
1719
1720 while( !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1721 || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1722 s->spatial_decomposition_count--;
1723
1724 if (s->spatial_decomposition_count <= 0) {
1725 av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1726 return AVERROR(EINVAL);
1727 }
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 < s->nb_planes; 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 < s->nb_planes; 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 FF_API_PRIVATE_OPT
1763 FF_DISABLE_DEPRECATION_WARNINGS
1764 if(s->avctx->scenechange_threshold)
1765 s->scenechange_threshold = s->avctx->scenechange_threshold;
1766 FF_ENABLE_DEPRECATION_WARNINGS
1767 #endif
1768
1769 if( plane_index==0
1770 && pic->pict_type == AV_PICTURE_TYPE_P
1771 && !(avctx->flags&AV_CODEC_FLAG_PASS2)
1772 && s->m.me.scene_change_score > s->scenechange_threshold){
1773 ff_init_range_encoder(c, pkt->data, pkt->size);
1774 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1775 pic->pict_type= AV_PICTURE_TYPE_I;
1776 s->keyframe=1;
1777 s->current_picture->key_frame=1;
1778 goto redo_frame;
1779 }
1780
1781 if(s->qlog == LOSSLESS_QLOG){
1782 for(y=0; y<h; y++){
1783 for(x=0; x<w; x++){
1784 s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1785 }
1786 }
1787 }else{
1788 for(y=0; y<h; y++){
1789 for(x=0; x<w; x++){
1790 s->spatial_dwt_buffer[y*w + x]= s->spatial_idwt_buffer[y*w + x] * (1 << ENCODER_EXTRA_BITS);
1791 }
1792 }
1793 }
1794
1795 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1796
1797 if(s->pass1_rc && plane_index==0){
1798 int delta_qlog = ratecontrol_1pass(s, pic);
1799 if (delta_qlog <= INT_MIN)
1800 return -1;
1801 if(delta_qlog){
1802 //reordering qlog in the bitstream would eliminate this reset
1803 ff_init_range_encoder(c, pkt->data, pkt->size);
1804 memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1805 memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1806 encode_header(s);
1807 encode_blocks(s, 0);
1808 }
1809 }
1810
1811 for(level=0; level<s->spatial_decomposition_count; level++){
1812 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1813 SubBand *b= &p->band[level][orientation];
1814
1815 quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1816 if(orientation==0)
1817 decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1818 if (!s->no_bitstream)
1819 encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1820 av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1821 if(orientation==0)
1822 correlate(s, b, b->ibuf, b->stride, 1, 0);
1823 }
1824 }
1825
1826 for(level=0; level<s->spatial_decomposition_count; level++){
1827 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1828 SubBand *b= &p->band[level][orientation];
1829
1830 dequantize(s, b, b->ibuf, b->stride);
1831 }
1832 }
1833
1834 ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1835 if(s->qlog == LOSSLESS_QLOG){
1836 for(y=0; y<h; y++){
1837 for(x=0; x<w; x++){
1838 s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1839 }
1840 }
1841 }
1842 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1843 }else{
1844 //ME/MC only
1845 if(pic->pict_type == AV_PICTURE_TYPE_I){
1846 for(y=0; y<h; y++){
1847 for(x=0; x<w; x++){
1848 s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1849 pict->data[plane_index][y*pict->linesize[plane_index] + x];
1850 }
1851 }
1852 }else{
1853 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1854 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1855 }
1856 }
1857 if(s->avctx->flags&AV_CODEC_FLAG_PSNR){
1858 int64_t error= 0;
1859
1860 if(pict->data[plane_index]) //FIXME gray hack
1861 for(y=0; y<h; y++){
1862 for(x=0; x<w; x++){
1863 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];
1864 error += d*d;
1865 }
1866 }
1867 s->avctx->error[plane_index] += error;
1868 s->encoding_error[plane_index] = error;
1869 }
1870
1871 }
1872 emms_c();
1873
1874 update_last_header_values(s);
1875
1876 ff_snow_release_buffer(avctx);
1877
1878 s->current_picture->coded_picture_number = avctx->frame_number;
1879 s->current_picture->pict_type = pic->pict_type;
1880 s->current_picture->quality = pic->quality;
1881 s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1882 s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1883 s->m.current_picture.f->display_picture_number =
1884 s->m.current_picture.f->coded_picture_number = avctx->frame_number;
1885 s->m.current_picture.f->quality = pic->quality;
1886 s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1887 if(s->pass1_rc)
1888 if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1889 return -1;
1890 if(avctx->flags&AV_CODEC_FLAG_PASS1)
1891 ff_write_pass1_stats(&s->m);
1892 s->m.last_pict_type = s->m.pict_type;
1893 #if FF_API_STAT_BITS
1894 FF_DISABLE_DEPRECATION_WARNINGS
1895 avctx->frame_bits = s->m.frame_bits;
1896 avctx->mv_bits = s->m.mv_bits;
1897 avctx->misc_bits = s->m.misc_bits;
1898 avctx->p_tex_bits = s->m.p_tex_bits;
1899 FF_ENABLE_DEPRECATION_WARNINGS
1900 #endif
1901
1902 emms_c();
1903
1904 ff_side_data_set_encoder_stats(pkt, s->current_picture->quality,
1905 s->encoding_error,
1906 (s->avctx->flags&AV_CODEC_FLAG_PSNR) ? 4 : 0,
1907 s->current_picture->pict_type);
1908
1909 #if FF_API_ERROR_FRAME
1910 FF_DISABLE_DEPRECATION_WARNINGS
1911 memcpy(s->current_picture->error, s->encoding_error, sizeof(s->encoding_error));
1912 FF_ENABLE_DEPRECATION_WARNINGS
1913 #endif
1914
1915 pkt->size = ff_rac_terminate(c, 0);
1916 if (s->current_picture->key_frame)
1917 pkt->flags |= AV_PKT_FLAG_KEY;
1918 *got_packet = 1;
1919
1920 return 0;
1921 }
1922
encode_end(AVCodecContext * avctx)1923 static av_cold int encode_end(AVCodecContext *avctx)
1924 {
1925 SnowContext *s = avctx->priv_data;
1926
1927 ff_snow_common_end(s);
1928 ff_rate_control_uninit(&s->m);
1929 av_frame_free(&s->input_picture);
1930 av_freep(&avctx->stats_out);
1931
1932 return 0;
1933 }
1934
1935 #define OFFSET(x) offsetof(SnowContext, x)
1936 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1937 static const AVOption options[] = {
1938 {"motion_est", "motion estimation algorithm", OFFSET(motion_est), AV_OPT_TYPE_INT, {.i64 = FF_ME_EPZS }, FF_ME_ZERO, FF_ME_ITER, VE, "motion_est" },
1939 { "zero", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ZERO }, 0, 0, VE, "motion_est" },
1940 { "epzs", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_EPZS }, 0, 0, VE, "motion_est" },
1941 { "xone", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_XONE }, 0, 0, VE, "motion_est" },
1942 { "iter", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ITER }, 0, 0, VE, "motion_est" },
1943 { "memc_only", "Only do ME/MC (I frames -> ref, P frame -> ME+MC).", OFFSET(memc_only), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1944 { "no_bitstream", "Skip final bitstream writeout.", OFFSET(no_bitstream), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1945 { "intra_penalty", "Penalty for intra blocks in block decission", OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1946 { "iterative_dia_size", "Dia size for the iterative ME", OFFSET(iterative_dia_size), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1947 { "sc_threshold", "Scene change threshold", OFFSET(scenechange_threshold), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, VE },
1948 { "pred", "Spatial decomposition type", OFFSET(pred), AV_OPT_TYPE_INT, { .i64 = 0 }, DWT_97, DWT_53, VE, "pred" },
1949 { "dwt97", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, VE, "pred" },
1950 { "dwt53", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, VE, "pred" },
1951 { "rc_eq", "Set rate control equation. When computing the expression, besides the standard functions "
1952 "defined in the section 'Expression Evaluation', the following functions are available: "
1953 "bits2qp(bits), qp2bits(qp). Also the following constants are available: iTex pTex tex mv "
1954 "fCode iCount mcVar var isI isP isB avgQP qComp avgIITex avgPITex avgPPTex avgBPTex avgTex.",
1955 OFFSET(m.rc_eq), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VE },
1956 { NULL },
1957 };
1958
1959 static const AVClass snowenc_class = {
1960 .class_name = "snow encoder",
1961 .item_name = av_default_item_name,
1962 .option = options,
1963 .version = LIBAVUTIL_VERSION_INT,
1964 };
1965
1966 AVCodec ff_snow_encoder = {
1967 .name = "snow",
1968 .long_name = NULL_IF_CONFIG_SMALL("Snow"),
1969 .type = AVMEDIA_TYPE_VIDEO,
1970 .id = AV_CODEC_ID_SNOW,
1971 .priv_data_size = sizeof(SnowContext),
1972 .init = encode_init,
1973 .encode2 = encode_frame,
1974 .close = encode_end,
1975 .pix_fmts = (const enum AVPixelFormat[]){
1976 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1977 AV_PIX_FMT_GRAY8,
1978 AV_PIX_FMT_NONE
1979 },
1980 .priv_class = &snowenc_class,
1981 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE |
1982 FF_CODEC_CAP_INIT_CLEANUP,
1983 };
1984