2 * Copyright (C) 2004 Michael Niedermayer <michaelni@gmx.at>
3 * Copyright (C) 2006 Robert Edele <yartrebo@earthlink.net>
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 #ifndef AVCODEC_SNOW_H
23 #define AVCODEC_SNOW_H
25 #include "libavutil/motion_vector.h"
32 #include "rangecoder.h"
35 #define FF_MPV_OFFSET(x) (offsetof(MpegEncContext, x) + offsetof(SnowContext, m))
36 #include "mpegvideo.h"
43 #define QROOT (1<<QSHIFT)
44 #define LOSSLESS_QLOG -128
46 #define MAX_REF_FRAMES 8
48 #define LOG2_OBMC_MAX 8
49 #define OBMC_MAX (1<<(LOG2_OBMC_MAX))
50 typedef struct BlockNode{
56 //#define TYPE_SPLIT 1
59 //#define TYPE_NOCOLOR 4
60 uint8_t level; //FIXME merge into type?
63 static const BlockNode null_block= { //FIXME add border maybe
64 .color= {128,128,128},
72 #define LOG2_MB_SIZE 4
73 #define MB_SIZE (1<<LOG2_MB_SIZE)
74 #define ENCODER_EXTRA_BITS 4
77 typedef struct x_and_coeff{
82 typedef struct SubBand{
87 int qlog; ///< log(qscale)/log[2^(1/6)]
92 int stride_line; ///< Stride measured in lines, not pixels.
93 x_and_coeff * x_coeff;
94 struct SubBand *parent;
95 uint8_t state[/*7*2*/ 7 + 512][32];
101 SubBand band[MAX_DECOMPOSITIONS][4];
104 int8_t hcoeff[HTAPS_MAX/2];
109 int8_t last_hcoeff[HTAPS_MAX/2];
113 typedef struct SnowContext{
115 AVCodecContext *avctx;
120 VideoDSPContext vdsp;
121 H264QpelContext h264qpel;
122 MpegvideoEncDSPContext mpvencdsp;
124 AVFrame *input_picture; ///< new_picture with the internal linesizes
125 AVFrame *current_picture;
126 AVFrame *last_picture[MAX_REF_FRAMES];
127 uint8_t *halfpel_plane[MAX_REF_FRAMES][4][4];
128 AVFrame *mconly_picture;
129 // uint8_t q_context[16];
130 uint8_t header_state[32];
131 uint8_t block_state[128 + 32*128];
135 int spatial_decomposition_type;
136 int last_spatial_decomposition_type;
137 int temporal_decomposition_type;
138 int spatial_decomposition_count;
139 int last_spatial_decomposition_count;
140 int temporal_decomposition_count;
143 int16_t (*ref_mvs[MAX_REF_FRAMES])[2];
144 uint32_t *ref_scores[MAX_REF_FRAMES];
145 DWTELEM *spatial_dwt_buffer;
146 DWTELEM *temp_dwt_buffer;
147 IDWTELEM *spatial_idwt_buffer;
148 IDWTELEM *temp_idwt_buffer;
153 int spatial_scalability;
163 #define QBIAS_SHIFT 3
167 int last_block_max_depth;
169 Plane plane[MAX_PLANES];
171 #define ME_CACHE_SIZE 1024
172 unsigned me_cache[ME_CACHE_SIZE];
173 unsigned me_cache_generation;
179 int iterative_dia_size;
181 MpegEncContext m; // needed for motion estimation, should not be used for anything else, the idea is to eventually make the motion estimation independent of MpegEncContext, so this will be removed then (FIXME/XXX)
184 uint8_t *emu_edge_buffer;
186 AVMotionVector *avmv;
188 uint64_t encoding_error[AV_NUM_DATA_POINTERS];
192 extern const uint8_t * const ff_obmc_tab[4];
193 extern uint8_t ff_qexp[QROOT];
194 extern int ff_scale_mv_ref[MAX_REF_FRAMES][MAX_REF_FRAMES];
196 /* C bits used by mmx/sse2/altivec */
198 static av_always_inline void snow_interleave_line_header(int * i, int width, IDWTELEM * low, IDWTELEM * high){
202 low[(*i)+1] = low[((*i)+1)>>1];
207 static av_always_inline void snow_interleave_line_footer(int * i, IDWTELEM * low, IDWTELEM * high){
208 for (; (*i)>=0; (*i)-=2){
209 low[(*i)+1] = high[(*i)>>1];
210 low[*i] = low[(*i)>>1];
214 static av_always_inline void snow_horizontal_compose_lift_lead_out(int i, IDWTELEM * dst, IDWTELEM * src, IDWTELEM * ref, int width, int w, int lift_high, int mul, int add, int shift){
216 dst[i] = src[i] - ((mul * (ref[i] + ref[i + 1]) + add) >> shift);
219 if((width^lift_high)&1){
220 dst[w] = src[w] - ((mul * 2 * ref[w] + add) >> shift);
224 static av_always_inline void snow_horizontal_compose_liftS_lead_out(int i, IDWTELEM * dst, IDWTELEM * src, IDWTELEM * ref, int width, int w){
226 dst[i] = src[i] + ((ref[i] + ref[(i+1)]+W_BO + 4 * src[i]) >> W_BS);
230 dst[w] = src[w] + ((2 * ref[w] + W_BO + 4 * src[w]) >> W_BS);
236 int ff_snow_common_init(AVCodecContext *avctx);
237 int ff_snow_common_init_after_header(AVCodecContext *avctx);
238 void ff_snow_common_end(SnowContext *s);
239 void ff_snow_release_buffer(AVCodecContext *avctx);
240 void ff_snow_reset_contexts(SnowContext *s);
241 int ff_snow_alloc_blocks(SnowContext *s);
242 int ff_snow_frame_start(SnowContext *s);
243 void ff_snow_pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, ptrdiff_t stride,
244 int sx, int sy, int b_w, int b_h, const BlockNode *block,
245 int plane_index, int w, int h);
246 int ff_snow_get_buffer(SnowContext *s, AVFrame *frame);
247 /* common inline functions */
248 //XXX doublecheck all of them should stay inlined
250 static inline void snow_set_blocks(SnowContext *s, int level, int x, int y, int l, int cb, int cr, int mx, int my, int ref, int type){
251 const int w= s->b_width << s->block_max_depth;
252 const int rem_depth= s->block_max_depth - level;
253 const int index= (x + y*w) << rem_depth;
254 const int block_w= 1<<rem_depth;
267 for(j=0; j<block_w; j++){
268 for(i=0; i<block_w; i++){
269 s->block[index + i + j*w]= block;
274 static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref,
275 const BlockNode *left, const BlockNode *top, const BlockNode *tr){
276 if(s->ref_frames == 1){
277 *mx = mid_pred(left->mx, top->mx, tr->mx);
278 *my = mid_pred(left->my, top->my, tr->my);
280 const int *scale = ff_scale_mv_ref[ref];
281 *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8,
282 (top ->mx * scale[top ->ref] + 128) >>8,
283 (tr ->mx * scale[tr ->ref] + 128) >>8);
284 *my = mid_pred((left->my * scale[left->ref] + 128) >>8,
285 (top ->my * scale[top ->ref] + 128) >>8,
286 (tr ->my * scale[tr ->ref] + 128) >>8);
290 static av_always_inline int same_block(BlockNode *a, BlockNode *b){
291 if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){
292 return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2]));
294 return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA));
298 //FIXME name cleanup (b_w, block_w, b_width stuff)
299 //XXX should we really inline it?
300 static av_always_inline void add_yblock(SnowContext *s, int sliced, slice_buffer *sb, IDWTELEM *dst, uint8_t *dst8, const uint8_t *obmc, int src_x, int src_y, int b_w, int b_h, int w, int h, int dst_stride, int src_stride, int obmc_stride, int b_x, int b_y, int add, int offset_dst, int plane_index){
301 const int b_width = s->b_width << s->block_max_depth;
302 const int b_height= s->b_height << s->block_max_depth;
303 const int b_stride= b_width;
304 BlockNode *lt= &s->block[b_x + b_y*b_stride];
306 BlockNode *lb= lt+b_stride;
309 // When src_stride is large enough, it is possible to interleave the blocks.
310 // Otherwise the blocks are written sequentially in the tmp buffer.
311 int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride;
312 uint8_t *tmp = s->scratchbuf;
319 }else if(b_x + 1 >= b_width){
326 }else if(b_y + 1 >= b_height){
331 if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
334 if(!sliced && !offset_dst)
342 obmc -= src_y*obmc_stride;
344 if(!sliced && !offset_dst)
345 dst -= src_y*dst_stride;
352 if(b_w<=0 || b_h<=0) return;
354 if(!sliced && offset_dst)
355 dst += src_x + src_y*dst_stride;
356 dst8+= src_x + src_y*src_stride;
357 // src += src_x + src_y*src_stride;
359 ptmp= tmp + 3*tmp_step;
362 ff_snow_pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h);
364 if(same_block(lt, rt)){
369 ff_snow_pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h);
372 if(same_block(lt, lb)){
374 }else if(same_block(rt, lb)){
379 ff_snow_pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h);
382 if(same_block(lt, rb) ){
384 }else if(same_block(rt, rb)){
386 }else if(same_block(lb, rb)){
390 ff_snow_pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h);
393 s->dwt.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
395 for(y=0; y<b_h; y++){
396 //FIXME ugly misuse of obmc_stride
397 const uint8_t *obmc1= obmc + y*obmc_stride;
398 const uint8_t *obmc2= obmc1+ (obmc_stride>>1);
399 const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
400 const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
401 for(x=0; x<b_w; x++){
402 int v= obmc1[x] * block[3][x + y*src_stride]
403 +obmc2[x] * block[2][x + y*src_stride]
404 +obmc3[x] * block[1][x + y*src_stride]
405 +obmc4[x] * block[0][x + y*src_stride];
407 v <<= 8 - LOG2_OBMC_MAX;
412 v += dst[x + y*dst_stride];
413 v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;
414 if(v&(~255)) v= ~(v>>31);
415 dst8[x + y*src_stride] = v;
417 dst[x + y*dst_stride] -= v;
424 static av_always_inline void predict_slice(SnowContext *s, IDWTELEM *buf, int plane_index, int add, int mb_y){
425 Plane *p= &s->plane[plane_index];
426 const int mb_w= s->b_width << s->block_max_depth;
427 const int mb_h= s->b_height << s->block_max_depth;
429 int block_size = MB_SIZE >> s->block_max_depth;
430 int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
431 int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
432 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
433 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
434 int ref_stride= s->current_picture->linesize[plane_index];
435 uint8_t *dst8= s->current_picture->data[plane_index];
438 av_assert2(s->chroma_h_shift == s->chroma_v_shift); // obmc params assume squares
439 if(s->keyframe || (s->avctx->debug&512)){
444 for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
446 int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
448 if(v&(~255)) v= ~(v>>31);
449 dst8[x + y*ref_stride]= v;
453 for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
455 buf[x + y*w]-= 128<<FRAC_BITS;
463 for(mb_x=0; mb_x<=mb_w; mb_x++){
464 add_yblock(s, 0, NULL, buf, dst8, obmc,
465 block_w*mb_x - block_w/2,
466 block_h*mb_y - block_h/2,
469 w, ref_stride, obmc_stride,
471 add, 1, plane_index);
475 static av_always_inline void predict_plane(SnowContext *s, IDWTELEM *buf, int plane_index, int add){
476 const int mb_h= s->b_height << s->block_max_depth;
478 for(mb_y=0; mb_y<=mb_h; mb_y++)
479 predict_slice(s, buf, plane_index, add, mb_y);
482 static inline void set_blocks(SnowContext *s, int level, int x, int y, int l, int cb, int cr, int mx, int my, int ref, int type){
483 const int w= s->b_width << s->block_max_depth;
484 const int rem_depth= s->block_max_depth - level;
485 const int index= (x + y*w) << rem_depth;
486 const int block_w= 1<<rem_depth;
487 const int block_h= 1<<rem_depth; //FIXME "w!=h"
500 for(j=0; j<block_h; j++){
501 for(i=0; i<block_w; i++){
502 s->block[index + i + j*w]= block;
507 static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
508 SnowContext *s = c->avctx->priv_data;
509 const int offset[3]= {
511 ((y*c->uvstride + x)>>s->chroma_h_shift),
512 ((y*c->uvstride + x)>>s->chroma_h_shift),
516 c->src[0][i]= src [i];
517 c->ref[0][i]= ref [i] + offset[i];
519 av_assert2(!ref_index);
523 /* bitstream functions */
525 extern const int8_t ff_quant3bA[256];
527 #define QEXPSHIFT (7-FRAC_BITS+8) //FIXME try to change this to 0
529 static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){
533 const int a= FFABS(v);
534 const int e= av_log2(a);
535 const int el= FFMIN(e, 10);
536 put_rac(c, state+0, 0);
539 put_rac(c, state+1+i, 1); //1..10
542 put_rac(c, state+1+9, 1); //1..10
544 put_rac(c, state+1+FFMIN(i,9), 0);
546 for(i=e-1; i>=el; i--){
547 put_rac(c, state+22+9, (a>>i)&1); //22..31
550 put_rac(c, state+22+i, (a>>i)&1); //22..31
554 put_rac(c, state+11 + el, v < 0); //11..21
556 put_rac(c, state+0, 1);
560 static inline int get_symbol(RangeCoder *c, uint8_t *state, int is_signed){
561 if(get_rac(c, state+0))
566 while(get_rac(c, state+1 + FFMIN(e,9))){ //1..10
569 return AVERROR_INVALIDDATA;
573 for(i=e-1; i>=0; i--){
574 a += a + get_rac(c, state+22 + FFMIN(i,9)); //22..31
577 e= -(is_signed && get_rac(c, state+11 + FFMIN(e,10))); //11..21
582 static inline void put_symbol2(RangeCoder *c, uint8_t *state, int v, int log2){
584 int r= log2>=0 ? 1<<log2 : 1;
587 av_assert2(log2>=-4);
590 put_rac(c, state+4+log2, 1);
595 put_rac(c, state+4+log2, 0);
597 for(i=log2-1; i>=0; i--){
598 put_rac(c, state+31-i, (v>>i)&1);
602 static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){
604 int r= log2>=0 ? 1<<log2 : 1;
607 av_assert2(log2>=-4);
609 while(log2<28 && get_rac(c, state+4+log2)){
615 for(i=log2-1; i>=0; i--){
616 v+= get_rac(c, state+31-i)<<i;
622 static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){
623 const int w= b->width;
624 const int h= b->height;
628 x_and_coeff *xc= b->x_coeff;
629 x_and_coeff *prev_xc= NULL;
630 x_and_coeff *prev2_xc= xc;
631 x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL;
632 x_and_coeff *prev_parent_xc= parent_xc;
634 runs= get_symbol2(&s->c, b->state[30], 0);
635 if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
642 if(y && prev_xc->x == 0){
654 if(prev_xc->x == x + 1)
660 if(x>>1 > parent_xc->x){
663 if(x>>1 == parent_xc->x){
667 if(/*ll|*/l|lt|t|rt|p){
668 int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1));
670 v=get_rac(&s->c, &b->state[0][context]);
672 v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1);
673 v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l&0xFF] + 3*ff_quant3bA[t&0xFF]]);
674 if ((uint16_t)v != v) {
675 av_log(s->avctx, AV_LOG_ERROR, "Coefficient damaged\n");
683 if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
685 v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1);
686 v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]);
687 if ((uint16_t)v != v) {
688 av_log(s->avctx, AV_LOG_ERROR, "Coefficient damaged\n");
698 av_assert2(run >= 0);
699 if(y) max_run= FFMIN(run, prev_xc->x - x - 2);
700 else max_run= FFMIN(run, w-x-1);
702 max_run= FFMIN(max_run, 2*parent_xc->x - x - 1);
703 av_assert2(max_run >= 0 && max_run <= run);
710 (xc++)->x= w+1; //end marker
716 while(parent_xc->x != parent->width+1)
719 prev_parent_xc= parent_xc;
721 parent_xc= prev_parent_xc;
726 (xc++)->x= w+1; //end marker
729 #endif /* AVCODEC_SNOW_H */