2 * The simplest mpeg encoder (well, it was the simplest!)
3 * Copyright (c) 2000,2001 Gerard Lantau.
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 * 4MV & hq & b-frame encoding stuff by Michael Niedermayer <michaelni@gmx.at>
27 #include "mpegvideo.h"
30 #include "fastmemcpy.h"
33 static void encode_picture(MpegEncContext *s, int picture_number);
34 static void rate_control_init(MpegEncContext *s);
35 static int rate_estimate_qscale(MpegEncContext *s);
36 static void dct_unquantize_mpeg1_c(MpegEncContext *s,
37 DCTELEM *block, int n, int qscale);
38 static void dct_unquantize_mpeg2_c(MpegEncContext *s,
39 DCTELEM *block, int n, int qscale);
40 static void dct_unquantize_h263_c(MpegEncContext *s,
41 DCTELEM *block, int n, int qscale);
42 static void draw_edges_c(UINT8 *buf, int wrap, int width, int height, int w);
43 static int dct_quantize_c(MpegEncContext *s, DCTELEM *block, int n, int qscale);
45 int (*dct_quantize)(MpegEncContext *s, DCTELEM *block, int n, int qscale)= dct_quantize_c;
46 void (*draw_edges)(UINT8 *buf, int wrap, int width, int height, int w)= draw_edges_c;
50 /* enable all paranoid tests for rounding, overflows, etc... */
56 /* for jpeg fast DCT */
59 static const unsigned short aanscales[64] = {
60 /* precomputed values scaled up by 14 bits */
61 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
62 22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270,
63 21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906,
64 19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315,
65 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
66 12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552,
67 8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446,
68 4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247
71 static UINT8 h263_chroma_roundtab[16] = {
72 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2,
75 static UINT16 default_mv_penalty[MAX_FCODE+1][MAX_MV*2+1];
76 static UINT8 default_fcode_tab[MAX_MV*2+1];
78 extern UINT8 zigzag_end[64];
80 /* default motion estimation */
81 int motion_estimation_method = ME_EPZS;
83 static void convert_matrix(int *qmat, UINT16 *qmat16, const UINT16 *quant_matrix, int qscale)
87 if (av_fdct == jpeg_fdct_ifast) {
89 /* 16 <= qscale * quant_matrix[i] <= 7905 */
90 /* 19952 <= aanscales[i] * qscale * quant_matrix[i] <= 249205026 */
91 /* (1<<36)/19952 >= (1<<36)/(aanscales[i] * qscale * quant_matrix[i]) >= (1<<36)/249205026 */
92 /* 3444240 >= (1<<36)/(aanscales[i] * qscale * quant_matrix[i]) >= 275 */
94 qmat[block_permute_op(i)] = (int)((UINT64_C(1) << (QMAT_SHIFT + 11)) /
95 (aanscales[i] * qscale * quant_matrix[block_permute_op(i)]));
99 /* We can safely suppose that 16 <= quant_matrix[i] <= 255
100 So 16 <= qscale * quant_matrix[i] <= 7905
101 so (1<<19) / 16 >= (1<<19) / (qscale * quant_matrix[i]) >= (1<<19) / 7905
102 so 32768 >= (1<<19) / (qscale * quant_matrix[i]) >= 67
104 qmat[i] = (1 << QMAT_SHIFT_MMX) / (qscale * quant_matrix[i]);
105 qmat16[i] = (1 << QMAT_SHIFT_MMX) / (qscale * quant_matrix[block_permute_op(i)]);
110 /* init common structure for both encoder and decoder */
111 int MPV_common_init(MpegEncContext *s)
116 s->dct_unquantize_h263 = dct_unquantize_h263_c;
117 s->dct_unquantize_mpeg1 = dct_unquantize_mpeg1_c;
118 s->dct_unquantize_mpeg2 = dct_unquantize_mpeg2_c;
121 MPV_common_init_mmx(s);
123 //setup default unquantizers (mpeg4 might change it later)
124 if(s->out_format == FMT_H263)
125 s->dct_unquantize = s->dct_unquantize_h263;
127 s->dct_unquantize = s->dct_unquantize_mpeg1;
129 s->mb_width = (s->width + 15) / 16;
130 s->mb_height = (s->height + 15) / 16;
131 s->mb_num = s->mb_width * s->mb_height;
132 s->linesize = s->mb_width * 16 + 2 * EDGE_WIDTH;
135 int w, h, shift, pict_start;
138 h = s->mb_height * 16 + 2 * EDGE_WIDTH;
139 shift = (i == 0) ? 0 : 1;
140 c_size = (w >> shift) * (h >> shift);
141 pict_start = (w >> shift) * (EDGE_WIDTH >> shift) + (EDGE_WIDTH >> shift);
143 pict = av_mallocz(c_size);
146 s->last_picture_base[i] = pict;
147 s->last_picture[i] = pict + pict_start;
149 pict = av_mallocz(c_size);
152 s->next_picture_base[i] = pict;
153 s->next_picture[i] = pict + pict_start;
155 if (s->has_b_frames) {
156 pict = av_mallocz(c_size);
159 s->aux_picture_base[i] = pict;
160 s->aux_picture[i] = pict + pict_start;
166 int mv_table_size= (s->mb_width+2)*(s->mb_height+2);
168 /* Allocate MB type table */
169 s->mb_type = av_mallocz(s->mb_num * sizeof(char));
170 if (s->mb_type == NULL) {
175 s->mb_var = av_mallocz(s->mb_num * sizeof(INT16));
176 if (s->mb_var == NULL) {
181 /* Allocate MV tables */
182 s->p_mv_table = av_mallocz(mv_table_size * 2 * sizeof(INT16));
183 if (s->p_mv_table == NULL) {
187 s->last_p_mv_table = av_mallocz(mv_table_size * 2 * sizeof(INT16));
188 if (s->last_p_mv_table == NULL) {
192 s->b_forw_mv_table = av_mallocz(mv_table_size * 2 * sizeof(INT16));
193 if (s->b_forw_mv_table == NULL) {
197 s->b_back_mv_table = av_mallocz(mv_table_size * 2 * sizeof(INT16));
198 if (s->b_back_mv_table == NULL) {
202 s->b_bidir_forw_mv_table = av_mallocz(mv_table_size * 2 * sizeof(INT16));
203 if (s->b_bidir_forw_mv_table == NULL) {
207 s->b_bidir_back_mv_table = av_mallocz(mv_table_size * 2 * sizeof(INT16));
208 if (s->b_bidir_back_mv_table == NULL) {
212 s->b_direct_forw_mv_table = av_mallocz(mv_table_size * 2 * sizeof(INT16));
213 if (s->b_direct_forw_mv_table == NULL) {
217 s->b_direct_back_mv_table = av_mallocz(mv_table_size * 2 * sizeof(INT16));
218 if (s->b_direct_back_mv_table == NULL) {
222 s->b_direct_mv_table = av_mallocz(mv_table_size * 2 * sizeof(INT16));
223 if (s->b_direct_mv_table == NULL) {
229 for(j=0; j<REORDER_BUFFER_SIZE; j++){
235 h = s->mb_height * 16;
236 shift = (i == 0) ? 0 : 1;
237 c_size = (w >> shift) * (h >> shift);
239 pict = av_mallocz(c_size);
242 s->picture_buffer[j][i] = pict;
248 if (s->out_format == FMT_H263 || s->encoding) {
251 size = (2 * s->mb_width + 2) * (2 * s->mb_height + 2);
252 s->motion_val = malloc(size * 2 * sizeof(INT16));
253 if (s->motion_val == NULL)
255 memset(s->motion_val, 0, size * 2 * sizeof(INT16));
258 if (s->h263_pred || s->h263_plus) {
259 int y_size, c_size, i, size;
263 y_size = (2 * s->mb_width + 2) * (2 * s->mb_height + 2);
264 c_size = (s->mb_width + 2) * (s->mb_height + 2);
265 size = y_size + 2 * c_size;
266 s->dc_val[0] = malloc(size * sizeof(INT16));
267 if (s->dc_val[0] == NULL)
269 s->dc_val[1] = s->dc_val[0] + y_size;
270 s->dc_val[2] = s->dc_val[1] + c_size;
272 s->dc_val[0][i] = 1024;
275 s->ac_val[0] = av_mallocz(size * sizeof(INT16) * 16);
276 if (s->ac_val[0] == NULL)
278 s->ac_val[1] = s->ac_val[0] + y_size;
279 s->ac_val[2] = s->ac_val[1] + c_size;
282 s->coded_block = av_mallocz(y_size);
286 /* which mb is a intra block */
287 s->mbintra_table = av_mallocz(s->mb_num);
288 if (!s->mbintra_table)
290 memset(s->mbintra_table, 1, s->mb_num);
292 /* default structure is frame */
293 s->picture_structure = PICT_FRAME;
295 /* init macroblock skip table */
296 s->mbskip_table = av_mallocz(s->mb_num);
297 if (!s->mbskip_table)
300 s->block= s->intra_block;
302 s->context_initialized = 1;
309 #define CHECK_FREE(p)\
315 /* init common structure for both encoder and decoder */
316 void MPV_common_end(MpegEncContext *s)
320 CHECK_FREE(s->mb_type);
321 CHECK_FREE(s->mb_var);
322 CHECK_FREE(s->p_mv_table);
323 CHECK_FREE(s->last_p_mv_table);
324 CHECK_FREE(s->b_forw_mv_table);
325 CHECK_FREE(s->b_back_mv_table);
326 CHECK_FREE(s->b_bidir_forw_mv_table);
327 CHECK_FREE(s->b_bidir_back_mv_table);
328 CHECK_FREE(s->b_direct_forw_mv_table);
329 CHECK_FREE(s->b_direct_back_mv_table);
330 CHECK_FREE(s->b_direct_mv_table);
331 CHECK_FREE(s->motion_val);
332 CHECK_FREE(s->dc_val[0]);
333 CHECK_FREE(s->ac_val[0]);
334 CHECK_FREE(s->coded_block);
335 CHECK_FREE(s->mbintra_table);
337 CHECK_FREE(s->mbskip_table);
340 CHECK_FREE(s->last_picture_base[i]);
341 CHECK_FREE(s->next_picture_base[i]);
342 CHECK_FREE(s->aux_picture_base[i]);
343 for(j=0; j<REORDER_BUFFER_SIZE; j++){
344 CHECK_FREE(s->picture_buffer[j][i]);
347 s->context_initialized = 0;
350 /* init video encoder */
351 int MPV_encode_init(AVCodecContext *avctx)
353 MpegEncContext *s = avctx->priv_data;
356 avctx->pix_fmt = PIX_FMT_YUV420P;
358 s->bit_rate = avctx->bit_rate;
359 s->bit_rate_tolerance = avctx->bit_rate_tolerance;
360 s->frame_rate = avctx->frame_rate;
361 s->width = avctx->width;
362 s->height = avctx->height;
363 s->gop_size = avctx->gop_size;
364 s->rtp_mode = avctx->rtp_mode;
365 s->rtp_payload_size = avctx->rtp_payload_size;
366 if (avctx->rtp_callback)
367 s->rtp_callback = avctx->rtp_callback;
368 s->qmin= avctx->qmin;
369 s->qmax= avctx->qmax;
370 s->max_qdiff= avctx->max_qdiff;
371 s->qcompress= avctx->qcompress;
372 s->qblur= avctx->qblur;
374 s->aspect_ratio_info= avctx->aspect_ratio_info;
375 s->flags= avctx->flags;
376 s->max_b_frames= avctx->max_b_frames;
378 if (s->gop_size <= 1) {
386 if (avctx->me_method == 0)
387 /* For compatibility */
388 s->me_method = motion_estimation_method;
390 s->me_method = avctx->me_method;
393 s->fixed_qscale = (avctx->flags & CODEC_FLAG_QSCALE);
395 switch(avctx->codec->id) {
396 case CODEC_ID_MPEG1VIDEO:
397 s->out_format = FMT_MPEG1;
400 s->out_format = FMT_MJPEG;
401 s->intra_only = 1; /* force intra only for jpeg */
402 s->mjpeg_write_tables = 1; /* write all tables */
403 s->mjpeg_vsample[0] = 2; /* set up default sampling factors */
404 s->mjpeg_vsample[1] = 1; /* the only currently supported values */
405 s->mjpeg_vsample[2] = 1;
406 s->mjpeg_hsample[0] = 2;
407 s->mjpeg_hsample[1] = 1;
408 s->mjpeg_hsample[2] = 1;
409 if (mjpeg_init(s) < 0)
413 if (h263_get_picture_format(s->width, s->height) == 7) {
414 printf("Input picture size isn't suitable for h263 codec! try h263+\n");
417 s->out_format = FMT_H263;
420 s->out_format = FMT_H263;
422 s->rtp_payload_size = 1200;
424 s->unrestricted_mv = 1;
426 /* These are just to be sure */
431 s->out_format = FMT_H263;
435 s->out_format = FMT_H263;
437 s->unrestricted_mv = 1;
438 s->has_b_frames= s->max_b_frames ? 1 : 0;
440 case CODEC_ID_MSMPEG4V1:
441 s->out_format = FMT_H263;
444 s->unrestricted_mv = 1;
445 s->msmpeg4_version= 1;
447 case CODEC_ID_MSMPEG4V2:
448 s->out_format = FMT_H263;
451 s->unrestricted_mv = 1;
452 s->msmpeg4_version= 2;
454 case CODEC_ID_MSMPEG4V3:
455 s->out_format = FMT_H263;
458 s->unrestricted_mv = 1;
459 s->msmpeg4_version= 3;
465 if((s->flags&CODEC_FLAG_4MV) && !(s->flags&CODEC_FLAG_HQ)){
466 printf("4MV is currently only supported in HQ mode\n");
470 { /* set up some save defaults, some codecs might override them later */
475 memset(default_mv_penalty, 0, sizeof(UINT16)*(MAX_FCODE+1)*(2*MAX_MV+1));
476 memset(default_fcode_tab , 0, sizeof(UINT8)*(2*MAX_MV+1));
478 for(i=-16; i<16; i++){
479 default_fcode_tab[i + MAX_MV]= 1;
483 s->mv_penalty= default_mv_penalty;
484 s->fcode_tab= default_fcode_tab;
486 if (s->out_format == FMT_H263)
488 else if (s->out_format == FMT_MPEG1)
489 mpeg1_encode_init(s);
491 /* dont use mv_penalty table for crap MV as it would be confused */
492 if (s->me_method < ME_EPZS) s->mv_penalty = default_mv_penalty;
497 if (MPV_common_init(s) < 0)
500 /* init default q matrix */
502 s->intra_matrix[i] = default_intra_matrix[i];
503 s->non_intra_matrix[i] = default_non_intra_matrix[i];
506 /* rate control init */
507 rate_control_init(s);
509 s->picture_number = 0;
510 s->picture_in_gop_number = 0;
511 s->fake_picture_number = 0;
512 /* motion detector init */
519 int MPV_encode_end(AVCodecContext *avctx)
521 MpegEncContext *s = avctx->priv_data;
527 if (s->out_format == FMT_MJPEG)
533 /* draw the edges of width 'w' of an image of size width, height */
534 static void draw_edges_c(UINT8 *buf, int wrap, int width, int height, int w)
536 UINT8 *ptr, *last_line;
539 last_line = buf + (height - 1) * wrap;
542 memcpy(buf - (i + 1) * wrap, buf, width);
543 memcpy(last_line + (i + 1) * wrap, last_line, width);
547 for(i=0;i<height;i++) {
548 memset(ptr - w, ptr[0], w);
549 memset(ptr + width, ptr[width-1], w);
554 memset(buf - (i + 1) * wrap - w, buf[0], w); /* top left */
555 memset(buf - (i + 1) * wrap + width, buf[width-1], w); /* top right */
556 memset(last_line + (i + 1) * wrap - w, last_line[0], w); /* top left */
557 memset(last_line + (i + 1) * wrap + width, last_line[width-1], w); /* top right */
561 /* generic function for encode/decode called before a frame is coded/decoded */
562 void MPV_frame_start(MpegEncContext *s)
568 if (s->pict_type == B_TYPE) {
570 s->current_picture[i] = s->aux_picture[i];
573 s->last_non_b_pict_type= s->pict_type;
575 /* swap next and last */
576 tmp = s->last_picture[i];
577 s->last_picture[i] = s->next_picture[i];
578 s->next_picture[i] = tmp;
579 s->current_picture[i] = tmp;
584 /* generic function for encode/decode called after a frame has been coded/decoded */
585 void MPV_frame_end(MpegEncContext *s)
587 /* draw edge for correct motion prediction if outside */
588 if (s->pict_type != B_TYPE && !s->intra_only) {
589 if(s->avctx==NULL || s->avctx->codec->id!=CODEC_ID_MPEG4 || s->divx_version==500){
590 draw_edges(s->current_picture[0], s->linesize, s->mb_width*16, s->mb_height*16, EDGE_WIDTH);
591 draw_edges(s->current_picture[1], s->linesize/2, s->mb_width*8, s->mb_height*8, EDGE_WIDTH/2);
592 draw_edges(s->current_picture[2], s->linesize/2, s->mb_width*8, s->mb_height*8, EDGE_WIDTH/2);
594 /* mpeg4? / opendivx / xvid */
595 draw_edges(s->current_picture[0], s->linesize, s->width, s->height, EDGE_WIDTH);
596 draw_edges(s->current_picture[1], s->linesize/2, s->width/2, s->height/2, EDGE_WIDTH/2);
597 draw_edges(s->current_picture[2], s->linesize/2, s->width/2, s->height/2, EDGE_WIDTH/2);
603 /* reorder input for encoding */
604 void reorder_input(MpegEncContext *s, AVPicture *pict)
608 if(s->max_b_frames > FF_MAX_B_FRAMES) s->max_b_frames= FF_MAX_B_FRAMES;
610 // delay= s->max_b_frames+1; (or 0 if no b frames cuz decoder diff)
612 for(j=0; j<REORDER_BUFFER_SIZE-1; j++){
613 s->coded_order[j]= s->coded_order[j+1];
615 s->coded_order[j].picture[0]= s->coded_order[j].picture[1]= s->coded_order[j].picture[2]= NULL; //catch uninitalized buffers
617 switch(s->input_pict_type){
622 index= s->max_b_frames - s->b_frames_since_non_b;
623 s->b_frames_since_non_b=0;
626 index= s->max_b_frames + 1;
627 s->b_frames_since_non_b++;
630 //printf("index:%d type:%d strides: %d %d\n", index, s->input_pict_type, pict->linesize[0], s->linesize);
631 if( (index==0 || (s->flags&CODEC_FLAG_INPUT_PRESERVED))
632 && pict->linesize[0] == s->linesize
633 && pict->linesize[1] == s->linesize>>1
634 && pict->linesize[2] == s->linesize>>1){
637 s->coded_order[index].picture[i]= pict->data[i];
642 uint8_t *src = pict->data[i];
644 int src_wrap = pict->linesize[i];
645 int dest_wrap = s->linesize;
649 if(index==0) dest= s->last_picture[i]+16; //is current_picture indeed but the switch hapens after reordering
650 else dest= s->picture_buffer[s->picture_buffer_index][i];
658 s->coded_order[index].picture[i]= dest;
660 memcpy(dest, src, w);
666 s->picture_buffer_index++;
667 if(s->picture_buffer_index >= REORDER_BUFFER_SIZE-1) s->picture_buffer_index=0;
670 s->coded_order[index].pict_type = s->input_pict_type;
671 s->coded_order[index].qscale = s->input_qscale;
672 s->coded_order[index].force_type= s->force_input_type;
673 s->coded_order[index].picture_in_gop_number= s->input_picture_in_gop_number;
674 s->coded_order[index].picture_number= s->input_picture_number;
677 s->new_picture[i]= s->coded_order[0].picture[i];
681 int MPV_encode_picture(AVCodecContext *avctx,
682 unsigned char *buf, int buf_size, void *data)
684 MpegEncContext *s = avctx->priv_data;
685 AVPicture *pict = data;
687 s->input_qscale = avctx->quality;
689 init_put_bits(&s->pb, buf, buf_size, NULL, NULL);
691 s->force_input_type= (avctx->flags&CODEC_FLAG_TYPE) ?
692 (avctx->key_frame ? I_TYPE : P_TYPE) : 0;
693 if (!s->intra_only) {
694 /* first picture of GOP is intra */
695 if (s->input_picture_in_gop_number % s->gop_size==0 || s->force_input_type==I_TYPE){
696 s->input_picture_in_gop_number=0;
697 s->input_pict_type = I_TYPE;
698 }else if(s->max_b_frames==0){
699 s->input_pict_type = P_TYPE;
701 if(s->b_frames_since_non_b < s->max_b_frames) //FIXME more IQ
702 s->input_pict_type = B_TYPE;
704 s->input_pict_type = P_TYPE;
707 s->input_pict_type = I_TYPE;
710 reorder_input(s, pict);
713 if(s->coded_order[0].picture[0]){
715 s->pict_type= s->coded_order[0].pict_type;
716 if (s->fixed_qscale) /* the ratecontrol needs the last qscale so we dont touch it for CBR */
717 s->qscale= s->coded_order[0].qscale;
718 s->force_type= s->coded_order[0].force_type;
719 s->picture_in_gop_number= s->coded_order[0].picture_in_gop_number;
720 s->picture_number= s->coded_order[0].picture_number;
724 encode_picture(s, s->picture_number);
725 avctx->key_frame = (s->pict_type == I_TYPE);
726 avctx->header_bits = s->header_bits;
727 avctx->mv_bits = s->mv_bits;
728 avctx->misc_bits = s->misc_bits;
729 avctx->i_tex_bits = s->i_tex_bits;
730 avctx->p_tex_bits = s->p_tex_bits;
731 avctx->i_count = s->i_count;
732 avctx->p_count = s->p_count;
733 avctx->skip_count = s->skip_count;
737 if (s->out_format == FMT_MJPEG)
738 mjpeg_picture_trailer(s);
740 avctx->quality = s->qscale;
743 s->input_picture_number++;
744 s->input_picture_in_gop_number++;
746 flush_put_bits(&s->pb);
747 s->last_frame_bits= s->frame_bits;
748 s->frame_bits = (pbBufPtr(&s->pb) - s->pb.buf) * 8;
749 s->total_bits += s->frame_bits;
750 avctx->frame_bits = s->frame_bits;
751 //printf("fcode: %d, type: %d, head: %d, mv: %d, misc: %d, frame: %d, itex: %d, ptex: %d\n",
752 //s->f_code, avctx->key_frame, s->header_bits, s->mv_bits, s->misc_bits, s->frame_bits, s->i_tex_bits, s->p_tex_bits);
754 if (avctx->get_psnr) {
755 /* At this point pict->data should have the original frame */
756 /* an s->current_picture should have the coded/decoded frame */
757 get_psnr(pict->data, s->current_picture,
758 pict->linesize, s->linesize, avctx);
759 // printf("%f\n", avctx->psnr_y);
761 return pbBufPtr(&s->pb) - s->pb.buf;
764 static inline int clip(int a, int amin, int amax)
774 static inline void gmc1_motion(MpegEncContext *s,
775 UINT8 *dest_y, UINT8 *dest_cb, UINT8 *dest_cr,
777 UINT8 **ref_picture, int src_offset,
781 int offset, src_x, src_y, linesize;
782 int motion_x, motion_y;
784 if(s->real_sprite_warping_points>1) printf("more than 1 warp point isnt supported\n");
785 motion_x= s->sprite_offset[0][0];
786 motion_y= s->sprite_offset[0][1];
787 src_x = s->mb_x * 16 + (motion_x >> (s->sprite_warping_accuracy+1));
788 src_y = s->mb_y * 16 + (motion_y >> (s->sprite_warping_accuracy+1));
789 motion_x<<=(3-s->sprite_warping_accuracy);
790 motion_y<<=(3-s->sprite_warping_accuracy);
791 src_x = clip(src_x, -16, s->width);
792 if (src_x == s->width)
794 src_y = clip(src_y, -16, s->height);
795 if (src_y == s->height)
798 linesize = s->linesize;
799 ptr = ref_picture[0] + (src_y * linesize) + src_x + src_offset;
802 gmc1(dest_y , ptr , linesize, h, motion_x&15, motion_y&15, s->no_rounding);
803 gmc1(dest_y+8, ptr+8, linesize, h, motion_x&15, motion_y&15, s->no_rounding);
805 motion_x= s->sprite_offset[1][0];
806 motion_y= s->sprite_offset[1][1];
807 src_x = s->mb_x * 8 + (motion_x >> (s->sprite_warping_accuracy+1));
808 src_y = s->mb_y * 8 + (motion_y >> (s->sprite_warping_accuracy+1));
809 motion_x<<=(3-s->sprite_warping_accuracy);
810 motion_y<<=(3-s->sprite_warping_accuracy);
811 src_x = clip(src_x, -8, s->width>>1);
812 if (src_x == s->width>>1)
814 src_y = clip(src_y, -8, s->height>>1);
815 if (src_y == s->height>>1)
818 offset = (src_y * linesize>>1) + src_x + (src_offset>>1);
819 ptr = ref_picture[1] + offset;
820 gmc1(dest_cb + (dest_offset>>1), ptr, linesize>>1, h>>1, motion_x&15, motion_y&15, s->no_rounding);
821 ptr = ref_picture[2] + offset;
822 gmc1(dest_cr + (dest_offset>>1), ptr, linesize>>1, h>>1, motion_x&15, motion_y&15, s->no_rounding);
827 /* apply one mpeg motion vector to the three components */
828 static inline void mpeg_motion(MpegEncContext *s,
829 UINT8 *dest_y, UINT8 *dest_cb, UINT8 *dest_cr,
831 UINT8 **ref_picture, int src_offset,
832 int field_based, op_pixels_func *pix_op,
833 int motion_x, int motion_y, int h)
836 int dxy, offset, mx, my, src_x, src_y, height, linesize;
837 if(s->quarter_sample)
843 dxy = ((motion_y & 1) << 1) | (motion_x & 1);
844 src_x = s->mb_x * 16 + (motion_x >> 1);
845 src_y = s->mb_y * (16 >> field_based) + (motion_y >> 1);
847 /* WARNING: do no forget half pels */
848 height = s->height >> field_based;
849 src_x = clip(src_x, -16, s->width);
850 if (src_x == s->width)
852 src_y = clip(src_y, -16, height);
855 linesize = s->linesize << field_based;
856 ptr = ref_picture[0] + (src_y * linesize) + (src_x) + src_offset;
857 dest_y += dest_offset;
858 pix_op[dxy](dest_y, ptr, linesize, h);
859 pix_op[dxy](dest_y + 8, ptr + 8, linesize, h);
861 if (s->out_format == FMT_H263) {
863 if ((motion_x & 3) != 0)
865 if ((motion_y & 3) != 0)
872 dxy = ((my & 1) << 1) | (mx & 1);
877 src_x = s->mb_x * 8 + mx;
878 src_y = s->mb_y * (8 >> field_based) + my;
879 src_x = clip(src_x, -8, s->width >> 1);
880 if (src_x == (s->width >> 1))
882 src_y = clip(src_y, -8, height >> 1);
883 if (src_y == (height >> 1))
886 offset = (src_y * (linesize >> 1)) + src_x + (src_offset >> 1);
887 ptr = ref_picture[1] + offset;
888 pix_op[dxy](dest_cb + (dest_offset >> 1), ptr, linesize >> 1, h >> 1);
889 ptr = ref_picture[2] + offset;
890 pix_op[dxy](dest_cr + (dest_offset >> 1), ptr, linesize >> 1, h >> 1);
893 static inline void qpel_motion(MpegEncContext *s,
894 UINT8 *dest_y, UINT8 *dest_cb, UINT8 *dest_cr,
896 UINT8 **ref_picture, int src_offset,
897 int field_based, op_pixels_func *pix_op,
898 qpel_mc_func *qpix_op,
899 int motion_x, int motion_y, int h)
902 int dxy, offset, mx, my, src_x, src_y, height, linesize;
904 dxy = ((motion_y & 3) << 2) | (motion_x & 3);
905 src_x = s->mb_x * 16 + (motion_x >> 2);
906 src_y = s->mb_y * (16 >> field_based) + (motion_y >> 2);
908 height = s->height >> field_based;
909 src_x = clip(src_x, -16, s->width);
910 if (src_x == s->width)
912 src_y = clip(src_y, -16, height);
915 linesize = s->linesize << field_based;
916 ptr = ref_picture[0] + (src_y * linesize) + src_x + src_offset;
917 dest_y += dest_offset;
918 //printf("%d %d %d\n", src_x, src_y, dxy);
919 qpix_op[dxy](dest_y , ptr , linesize, linesize, motion_x&3, motion_y&3);
920 qpix_op[dxy](dest_y + 8, ptr + 8, linesize, linesize, motion_x&3, motion_y&3);
921 qpix_op[dxy](dest_y + linesize*8 , ptr + linesize*8 , linesize, linesize, motion_x&3, motion_y&3);
922 qpix_op[dxy](dest_y + linesize*8 + 8, ptr + linesize*8 + 8, linesize, linesize, motion_x&3, motion_y&3);
924 mx= (motion_x>>1) | (motion_x&1);
925 my= (motion_y>>1) | (motion_y&1);
935 src_x = s->mb_x * 8 + mx;
936 src_y = s->mb_y * (8 >> field_based) + my;
937 src_x = clip(src_x, -8, s->width >> 1);
938 if (src_x == (s->width >> 1))
940 src_y = clip(src_y, -8, height >> 1);
941 if (src_y == (height >> 1))
944 offset = (src_y * (linesize >> 1)) + src_x + (src_offset >> 1);
945 ptr = ref_picture[1] + offset;
946 pix_op[dxy](dest_cb + (dest_offset >> 1), ptr, linesize >> 1, h >> 1);
947 ptr = ref_picture[2] + offset;
948 pix_op[dxy](dest_cr + (dest_offset >> 1), ptr, linesize >> 1, h >> 1);
952 static inline void MPV_motion(MpegEncContext *s,
953 UINT8 *dest_y, UINT8 *dest_cb, UINT8 *dest_cr,
954 int dir, UINT8 **ref_picture,
955 op_pixels_func *pix_op, qpel_mc_func *qpix_op)
957 int dxy, offset, mx, my, src_x, src_y, motion_x, motion_y;
968 mpeg_motion(s, dest_y, dest_cb, dest_cr, 0,
971 s->sprite_offset[0][0]>>3,
972 s->sprite_offset[0][1]>>3,
975 gmc1_motion(s, dest_y, dest_cb, dest_cr, 0,
979 }else if(s->quarter_sample && dir==0){ //FIXME
980 qpel_motion(s, dest_y, dest_cb, dest_cr, 0,
983 s->mv[dir][0][0], s->mv[dir][0][1], 16);
985 mpeg_motion(s, dest_y, dest_cb, dest_cr, 0,
988 s->mv[dir][0][0], s->mv[dir][0][1], 16);
993 motion_x = s->mv[dir][i][0];
994 motion_y = s->mv[dir][i][1];
996 dxy = ((motion_y & 1) << 1) | (motion_x & 1);
997 src_x = mb_x * 16 + (motion_x >> 1) + (i & 1) * 8;
998 src_y = mb_y * 16 + (motion_y >> 1) + (i >>1) * 8;
1000 /* WARNING: do no forget half pels */
1001 src_x = clip(src_x, -16, s->width);
1002 if (src_x == s->width)
1004 src_y = clip(src_y, -16, s->height);
1005 if (src_y == s->height)
1008 ptr = ref_picture[0] + (src_y * s->linesize) + (src_x);
1009 dest = dest_y + ((i & 1) * 8) + (i >> 1) * 8 * s->linesize;
1010 pix_op[dxy](dest, ptr, s->linesize, 8);
1012 /* In case of 8X8, we construct a single chroma motion vector
1013 with a special rounding */
1017 mx += s->mv[dir][i][0];
1018 my += s->mv[dir][i][1];
1021 mx = (h263_chroma_roundtab[mx & 0xf] + ((mx >> 3) & ~1));
1024 mx = -(h263_chroma_roundtab[mx & 0xf] + ((mx >> 3) & ~1));
1027 my = (h263_chroma_roundtab[my & 0xf] + ((my >> 3) & ~1));
1030 my = -(h263_chroma_roundtab[my & 0xf] + ((my >> 3) & ~1));
1032 dxy = ((my & 1) << 1) | (mx & 1);
1036 src_x = mb_x * 8 + mx;
1037 src_y = mb_y * 8 + my;
1038 src_x = clip(src_x, -8, s->width/2);
1039 if (src_x == s->width/2)
1041 src_y = clip(src_y, -8, s->height/2);
1042 if (src_y == s->height/2)
1045 offset = (src_y * (s->linesize >> 1)) + src_x;
1046 ptr = ref_picture[1] + offset;
1047 pix_op[dxy](dest_cb, ptr, s->linesize >> 1, 8);
1048 ptr = ref_picture[2] + offset;
1049 pix_op[dxy](dest_cr, ptr, s->linesize >> 1, 8);
1052 if (s->picture_structure == PICT_FRAME) {
1054 mpeg_motion(s, dest_y, dest_cb, dest_cr, 0,
1055 ref_picture, s->field_select[dir][0] ? s->linesize : 0,
1057 s->mv[dir][0][0], s->mv[dir][0][1], 8);
1059 mpeg_motion(s, dest_y, dest_cb, dest_cr, s->linesize,
1060 ref_picture, s->field_select[dir][1] ? s->linesize : 0,
1062 s->mv[dir][1][0], s->mv[dir][1][1], 8);
1072 /* put block[] to dest[] */
1073 static inline void put_dct(MpegEncContext *s,
1074 DCTELEM *block, int i, UINT8 *dest, int line_size)
1077 s->dct_unquantize(s, block, i, s->qscale);
1079 put_pixels_clamped(block, dest, line_size);
1082 /* add block[] to dest[] */
1083 static inline void add_dct(MpegEncContext *s,
1084 DCTELEM *block, int i, UINT8 *dest, int line_size)
1086 if (s->block_last_index[i] >= 0) {
1088 if(s->encoding || (!s->h263_msmpeg4))
1089 s->dct_unquantize(s, block, i, s->qscale);
1092 add_pixels_clamped(block, dest, line_size);
1096 /* generic function called after a macroblock has been parsed by the
1097 decoder or after it has been encoded by the encoder.
1099 Important variables used:
1100 s->mb_intra : true if intra macroblock
1101 s->mv_dir : motion vector direction
1102 s->mv_type : motion vector type
1103 s->mv : motion vector
1104 s->interlaced_dct : true if interlaced dct used (mpeg2)
1106 void MPV_decode_mb(MpegEncContext *s, DCTELEM block[6][64])
1109 int dct_linesize, dct_offset;
1110 op_pixels_func *op_pix;
1111 qpel_mc_func *op_qpix;
1116 #ifdef FF_POSTPROCESS
1117 quant_store[mb_y][mb_x]=s->qscale;
1118 //printf("[%02d][%02d] %d\n",mb_x,mb_y,s->qscale);
1121 /* update DC predictors for P macroblocks */
1123 if (s->h263_pred || s->h263_aic) {
1124 if(s->mbintra_table[mb_x + mb_y*s->mb_width])
1127 s->mbintra_table[mb_x + mb_y*s->mb_width]=0;
1128 wrap = 2 * s->mb_width + 2;
1129 xy = 2 * mb_x + 1 + (2 * mb_y + 1) * wrap;
1132 s->dc_val[0][xy] = v;
1133 s->dc_val[0][xy + 1] = v;
1134 s->dc_val[0][xy + wrap] = v;
1135 s->dc_val[0][xy + 1 + wrap] = v;
1137 memset(s->ac_val[0][xy], 0, 16 * sizeof(INT16));
1138 memset(s->ac_val[0][xy + 1], 0, 16 * sizeof(INT16));
1139 memset(s->ac_val[0][xy + wrap], 0, 16 * sizeof(INT16));
1140 memset(s->ac_val[0][xy + 1 + wrap], 0, 16 * sizeof(INT16));
1141 if (s->h263_msmpeg4) {
1142 s->coded_block[xy] = 0;
1143 s->coded_block[xy + 1] = 0;
1144 s->coded_block[xy + wrap] = 0;
1145 s->coded_block[xy + 1 + wrap] = 0;
1148 wrap = s->mb_width + 2;
1149 xy = mb_x + 1 + (mb_y + 1) * wrap;
1150 s->dc_val[1][xy] = v;
1151 s->dc_val[2][xy] = v;
1153 memset(s->ac_val[1][xy], 0, 16 * sizeof(INT16));
1154 memset(s->ac_val[2][xy], 0, 16 * sizeof(INT16));
1157 s->last_dc[0] = 128 << s->intra_dc_precision;
1158 s->last_dc[1] = 128 << s->intra_dc_precision;
1159 s->last_dc[2] = 128 << s->intra_dc_precision;
1162 else if (s->h263_pred || s->h263_aic)
1163 s->mbintra_table[mb_x + mb_y*s->mb_width]=1;
1165 /* update motion predictor, not for B-frames as they need the motion_val from the last P/S-Frame */
1166 if (s->out_format == FMT_H263) { //FIXME move into h263.c if possible, format specific stuff shouldnt be here
1167 if(s->pict_type!=B_TYPE){
1168 int xy, wrap, motion_x, motion_y;
1170 wrap = 2 * s->mb_width + 2;
1171 xy = 2 * mb_x + 1 + (2 * mb_y + 1) * wrap;
1176 } else if (s->mv_type == MV_TYPE_16X16) {
1177 motion_x = s->mv[0][0][0];
1178 motion_y = s->mv[0][0][1];
1180 /* no update if 8X8 because it has been done during parsing */
1181 s->motion_val[xy][0] = motion_x;
1182 s->motion_val[xy][1] = motion_y;
1183 s->motion_val[xy + 1][0] = motion_x;
1184 s->motion_val[xy + 1][1] = motion_y;
1185 s->motion_val[xy + wrap][0] = motion_x;
1186 s->motion_val[xy + wrap][1] = motion_y;
1187 s->motion_val[xy + 1 + wrap][0] = motion_x;
1188 s->motion_val[xy + 1 + wrap][1] = motion_y;
1193 if (!(s->encoding && (s->intra_only || s->pict_type==B_TYPE))) {
1194 UINT8 *dest_y, *dest_cb, *dest_cr;
1197 /* avoid copy if macroblock skipped in last frame too
1198 dont touch it for B-frames as they need the skip info from the next p-frame */
1199 if (s->pict_type != B_TYPE) {
1200 mbskip_ptr = &s->mbskip_table[s->mb_y * s->mb_width + s->mb_x];
1203 /* if previous was skipped too, then nothing to do !
1204 skip only during decoding as we might trash the buffers during encoding a bit */
1205 if (*mbskip_ptr != 0 && !s->encoding)
1207 *mbskip_ptr = 1; /* indicate that this time we skiped it */
1209 *mbskip_ptr = 0; /* not skipped */
1213 dest_y = s->current_picture[0] + (mb_y * 16 * s->linesize) + mb_x * 16;
1214 dest_cb = s->current_picture[1] + (mb_y * 8 * (s->linesize >> 1)) + mb_x * 8;
1215 dest_cr = s->current_picture[2] + (mb_y * 8 * (s->linesize >> 1)) + mb_x * 8;
1217 if (s->interlaced_dct) {
1218 dct_linesize = s->linesize * 2;
1219 dct_offset = s->linesize;
1221 dct_linesize = s->linesize;
1222 dct_offset = s->linesize * 8;
1226 /* motion handling */
1227 if((s->flags&CODEC_FLAG_HQ) || (!s->encoding)){
1228 if (!s->no_rounding){
1229 op_pix = put_pixels_tab;
1230 op_qpix= qpel_mc_rnd_tab;
1232 op_pix = put_no_rnd_pixels_tab;
1233 op_qpix= qpel_mc_no_rnd_tab;
1236 if (s->mv_dir & MV_DIR_FORWARD) {
1237 MPV_motion(s, dest_y, dest_cb, dest_cr, 0, s->last_picture, op_pix, op_qpix);
1238 if (!s->no_rounding)
1239 op_pix = avg_pixels_tab;
1241 op_pix = avg_no_rnd_pixels_tab;
1243 if (s->mv_dir & MV_DIR_BACKWARD) {
1244 MPV_motion(s, dest_y, dest_cb, dest_cr, 1, s->next_picture, op_pix, op_qpix);
1248 /* add dct residue */
1249 add_dct(s, block[0], 0, dest_y, dct_linesize);
1250 add_dct(s, block[1], 1, dest_y + 8, dct_linesize);
1251 add_dct(s, block[2], 2, dest_y + dct_offset, dct_linesize);
1252 add_dct(s, block[3], 3, dest_y + dct_offset + 8, dct_linesize);
1254 add_dct(s, block[4], 4, dest_cb, s->linesize >> 1);
1255 add_dct(s, block[5], 5, dest_cr, s->linesize >> 1);
1257 /* dct only in intra block */
1258 put_dct(s, block[0], 0, dest_y, dct_linesize);
1259 put_dct(s, block[1], 1, dest_y + 8, dct_linesize);
1260 put_dct(s, block[2], 2, dest_y + dct_offset, dct_linesize);
1261 put_dct(s, block[3], 3, dest_y + dct_offset + 8, dct_linesize);
1263 put_dct(s, block[4], 4, dest_cb, s->linesize >> 1);
1264 put_dct(s, block[5], 5, dest_cr, s->linesize >> 1);
1268 emms_c(); //FIXME remove
1272 static void encode_mb(MpegEncContext *s, int motion_x, int motion_y)
1274 const int mb_x= s->mb_x;
1275 const int mb_y= s->mb_y;
1278 if (s->interlaced_dct) {
1279 dct_linesize = s->linesize * 2;
1280 dct_offset = s->linesize;
1282 dct_linesize = s->linesize;
1283 dct_offset = s->linesize * 8;
1292 ptr = s->new_picture[0] + (mb_y * 16 * wrap) + mb_x * 16;
1293 get_pixels(s->block[0], ptr , wrap);
1294 get_pixels(s->block[1], ptr + 8, wrap);
1295 get_pixels(s->block[2], ptr + 8 * wrap , wrap);
1296 get_pixels(s->block[3], ptr + 8 * wrap + 8, wrap);
1299 ptr = s->new_picture[1] + (mb_y * 8 * wrap) + mb_x * 8;
1300 get_pixels(s->block[4], ptr, wrap);
1302 ptr = s->new_picture[2] + (mb_y * 8 * wrap) + mb_x * 8;
1303 get_pixels(s->block[5], ptr, wrap);
1305 op_pixels_func *op_pix;
1306 qpel_mc_func *op_qpix;
1307 UINT8 *dest_y, *dest_cb, *dest_cr;
1311 dest_y = s->current_picture[0] + (mb_y * 16 * s->linesize ) + mb_x * 16;
1312 dest_cb = s->current_picture[1] + (mb_y * 8 * (s->linesize >> 1)) + mb_x * 8;
1313 dest_cr = s->current_picture[2] + (mb_y * 8 * (s->linesize >> 1)) + mb_x * 8;
1315 if (!s->no_rounding){
1316 op_pix = put_pixels_tab;
1317 op_qpix= qpel_mc_rnd_tab;
1319 op_pix = put_no_rnd_pixels_tab;
1320 op_qpix= qpel_mc_no_rnd_tab;
1323 if (s->mv_dir & MV_DIR_FORWARD) {
1324 MPV_motion(s, dest_y, dest_cb, dest_cr, 0, s->last_picture, op_pix, op_qpix);
1325 if (!s->no_rounding)
1326 op_pix = avg_pixels_tab;
1328 op_pix = avg_no_rnd_pixels_tab;
1330 if (s->mv_dir & MV_DIR_BACKWARD) {
1331 MPV_motion(s, dest_y, dest_cb, dest_cr, 1, s->next_picture, op_pix, op_qpix);
1334 ptr = s->new_picture[0] + (mb_y * 16 * wrap) + mb_x * 16;
1335 diff_pixels(s->block[0], ptr , dest_y , wrap);
1336 diff_pixels(s->block[1], ptr + 8, dest_y + 8, wrap);
1337 diff_pixels(s->block[2], ptr + 8 * wrap , dest_y + 8 * wrap , wrap);
1338 diff_pixels(s->block[3], ptr + 8 * wrap + 8, dest_y + 8 * wrap + 8, wrap);
1341 ptr = s->new_picture[1] + (mb_y * 8 * wrap) + mb_x * 8;
1342 diff_pixels(s->block[4], ptr, dest_cb, wrap);
1344 ptr = s->new_picture[2] + (mb_y * 8 * wrap) + mb_x * 8;
1345 diff_pixels(s->block[5], ptr, dest_cr, wrap);
1352 adap_parm = ((s->avg_mb_var << 1) + s->mb_var[s->mb_width*mb_y+mb_x] + 1.0) /
1353 ((s->mb_var[s->mb_width*mb_y+mb_x] << 1) + s->avg_mb_var + 1.0);
1355 printf("\ntype=%c qscale=%2d adap=%0.2f dquant=%4.2f var=%4d avgvar=%4d",
1356 (s->mb_type[s->mb_width*mb_y+mb_x] > 0) ? 'I' : 'P',
1357 s->qscale, adap_parm, s->qscale*adap_parm,
1358 s->mb_var[s->mb_width*mb_y+mb_x], s->avg_mb_var);
1361 /* DCT & quantize */
1362 if (s->h263_msmpeg4) {
1363 msmpeg4_dc_scale(s);
1364 } else if (s->h263_pred) {
1367 /* default quantization values */
1372 s->block_last_index[i] = dct_quantize(s, s->block[i], i, s->qscale);
1375 /* huffman encode */
1376 switch(s->out_format) {
1378 mpeg1_encode_mb(s, s->block, motion_x, motion_y);
1381 if (s->h263_msmpeg4)
1382 msmpeg4_encode_mb(s, s->block, motion_x, motion_y);
1383 else if(s->h263_pred)
1384 mpeg4_encode_mb(s, s->block, motion_x, motion_y);
1386 h263_encode_mb(s, s->block, motion_x, motion_y);
1389 mjpeg_encode_mb(s, s->block);
1394 static void copy_bits(PutBitContext *pb, UINT8 *src, int length)
1397 int bytes= length>>4;
1398 int bits= length&15;
1401 for(i=0; i<bytes; i++) put_bits(pb, 16, be2me_16(((uint16_t*)src)[i]));
1402 put_bits(pb, bits, be2me_16(((uint16_t*)src)[i])>>(16-bits));
1404 int bytes= length>>3;
1408 for(i=0; i<bytes; i++) put_bits(pb, 8, src[i]);
1409 put_bits(pb, bits, src[i]>>(8-bits));
1413 static void copy_context_before_encode(MpegEncContext *d, MpegEncContext *s, int type){
1416 memcpy(d->last_mv, s->last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster then a loop?
1419 d->mb_incr= s->mb_incr;
1421 d->last_dc[i]= s->last_dc[i];
1424 d->mv_bits= s->mv_bits;
1425 d->i_tex_bits= s->i_tex_bits;
1426 d->p_tex_bits= s->p_tex_bits;
1427 d->i_count= s->i_count;
1428 d->p_count= s->p_count;
1429 d->skip_count= s->skip_count;
1430 d->misc_bits= s->misc_bits;
1431 d->last_bits= s->last_bits;
1434 static void copy_context_after_encode(MpegEncContext *d, MpegEncContext *s, int type){
1437 memcpy(d->mv, s->mv, 2*4*2*sizeof(int));
1438 memcpy(d->last_mv, s->last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster then a loop?
1441 d->mb_incr= s->mb_incr;
1443 d->last_dc[i]= s->last_dc[i];
1446 d->mv_bits= s->mv_bits;
1447 d->i_tex_bits= s->i_tex_bits;
1448 d->p_tex_bits= s->p_tex_bits;
1449 d->i_count= s->i_count;
1450 d->p_count= s->p_count;
1451 d->skip_count= s->skip_count;
1452 d->misc_bits= s->misc_bits;
1453 d->last_bits= s->last_bits;
1455 d->mb_intra= s->mb_intra;
1456 d->mv_type= s->mv_type;
1457 d->mv_dir= s->mv_dir;
1461 d->block_last_index[i]= s->block_last_index[i];
1465 static void encode_picture(MpegEncContext *s, int picture_number)
1467 int mb_x, mb_y, last_gob, pdif = 0;
1470 MpegEncContext best_s, backup_s;
1471 UINT8 bit_buf[4][3000]; //FIXME check that this is ALLWAYS large enogh for a MB
1473 s->picture_number = picture_number;
1478 s->block_wrap[3]= s->mb_width*2 + 2;
1480 s->block_wrap[5]= s->mb_width + 2;
1482 s->last_mc_mb_var = s->mc_mb_var;
1483 /* Reset the average MB variance */
1487 /* Estimate motion for every MB */
1488 if(s->pict_type != I_TYPE){
1489 // int16_t (*tmp)[2]= s->p_mv_table;
1490 // s->p_mv_table= s->last_mv_table;
1491 // s->last_mv_table= s->mv_table;
1493 for(mb_y=0; mb_y < s->mb_height; mb_y++) {
1494 s->block_index[0]= s->block_wrap[0]*(mb_y*2 + 1) - 1;
1495 s->block_index[1]= s->block_wrap[0]*(mb_y*2 + 1);
1496 s->block_index[2]= s->block_wrap[0]*(mb_y*2 + 2) - 1;
1497 s->block_index[3]= s->block_wrap[0]*(mb_y*2 + 2);
1498 for(mb_x=0; mb_x < s->mb_width; mb_x++) {
1501 s->block_index[0]+=2;
1502 s->block_index[1]+=2;
1503 s->block_index[2]+=2;
1504 s->block_index[3]+=2;
1506 /* compute motion vector & mb_type and store in context */
1507 if(s->pict_type==B_TYPE)
1508 ff_estimate_b_frame_motion(s, mb_x, mb_y);
1510 ff_estimate_p_frame_motion(s, mb_x, mb_y);
1511 // s->mb_type[mb_y*s->mb_width + mb_x]=MB_TYPE_INTER;
1515 }else if(s->pict_type == I_TYPE){
1517 //FIXME do we need to zero them?
1518 memset(s->motion_val[0], 0, sizeof(INT16)*(s->mb_width*2 + 2)*(s->mb_height*2 + 2)*2);
1519 memset(s->p_mv_table , 0, sizeof(INT16)*(s->mb_width+2)*(s->mb_height+2)*2);
1520 memset(s->mb_type , MB_TYPE_INTRA, sizeof(UINT8)*s->mb_width*s->mb_height);
1523 if(s->avg_mb_var < s->mc_mb_var && s->pict_type != B_TYPE && (!s->force_type) && s->max_b_frames==0){ //FIXME subtract MV bits
1524 // FIXME b-frames & scene change detection
1525 s->input_pict_type= I_TYPE;
1526 s->pict_type= I_TYPE;
1527 s->input_picture_in_gop_number=0;
1528 memset(s->mb_type , MB_TYPE_INTRA, sizeof(UINT8)*s->mb_width*s->mb_height);
1529 //printf("Scene change detected, encoding as I Frame\n");
1532 if(s->pict_type==P_TYPE || s->pict_type==S_TYPE)
1533 s->f_code= ff_get_best_fcode(s, s->p_mv_table, MB_TYPE_INTER);
1534 ff_fix_long_p_mvs(s);
1535 if(s->pict_type==B_TYPE){
1536 s->f_code= ff_get_best_fcode(s, s->b_forw_mv_table, MB_TYPE_FORWARD);
1537 s->b_code= ff_get_best_fcode(s, s->b_back_mv_table, MB_TYPE_BACKWARD);
1538 //FIXME if BIDIR != for&back
1539 ff_fix_long_b_mvs(s, s->b_forw_mv_table, s->f_code, MB_TYPE_FORWARD |MB_TYPE_BIDIR);
1540 ff_fix_long_b_mvs(s, s->b_back_mv_table, s->b_code, MB_TYPE_BACKWARD|MB_TYPE_BIDIR);
1543 //printf("f_code %d ///\n", s->f_code);
1545 // printf("%d %d\n", s->avg_mb_var, s->mc_mb_var);
1547 if (!s->fixed_qscale)
1548 s->qscale = rate_estimate_qscale(s);
1550 /* precompute matrix */
1551 if (s->out_format == FMT_MJPEG) {
1552 /* for mjpeg, we do include qscale in the matrix */
1553 s->intra_matrix[0] = default_intra_matrix[0];
1555 s->intra_matrix[i] = (default_intra_matrix[i] * s->qscale) >> 3;
1556 convert_matrix(s->q_intra_matrix, s->q_intra_matrix16, s->intra_matrix, 8);
1558 convert_matrix(s->q_intra_matrix, s->q_intra_matrix16, s->intra_matrix, s->qscale);
1559 convert_matrix(s->q_non_intra_matrix, s->q_non_intra_matrix16, s->non_intra_matrix, s->qscale);
1562 s->last_bits= get_bit_count(&s->pb);
1563 switch(s->out_format) {
1565 mjpeg_picture_header(s);
1568 if (s->h263_msmpeg4)
1569 msmpeg4_encode_picture_header(s, picture_number);
1570 else if (s->h263_pred)
1571 mpeg4_encode_picture_header(s, picture_number);
1572 else if (s->h263_rv10)
1573 rv10_encode_picture_header(s, picture_number);
1575 h263_encode_picture_header(s, picture_number);
1578 mpeg1_encode_picture_header(s, picture_number);
1581 bits= get_bit_count(&s->pb);
1582 s->header_bits= bits - s->last_bits;
1592 /* init last dc values */
1593 /* note: quant matrix value (8) is implied here */
1594 s->last_dc[0] = 128;
1595 s->last_dc[1] = 128;
1596 s->last_dc[2] = 128;
1598 s->last_mv[0][0][0] = 0;
1599 s->last_mv[0][0][1] = 0;
1601 /* Get the GOB height based on picture height */
1602 if (s->out_format == FMT_H263 && !s->h263_pred && !s->h263_msmpeg4) {
1603 if (s->height <= 400)
1605 else if (s->height <= 800)
1611 s->avg_mb_var = s->avg_mb_var / s->mb_num;
1613 for(mb_y=0; mb_y < s->mb_height; mb_y++) {
1614 /* Put GOB header based on RTP MTU */
1615 /* TODO: Put all this stuff in a separate generic function */
1618 s->ptr_lastgob = s->pb.buf;
1619 s->ptr_last_mb_line = s->pb.buf;
1620 } else if (s->out_format == FMT_H263 && !s->h263_pred && !s->h263_msmpeg4 && !(mb_y % s->gob_index)) {
1621 last_gob = h263_encode_gob_header(s, mb_y);
1623 s->first_gob_line = 1;
1628 s->block_index[0]= s->block_wrap[0]*(mb_y*2 + 1) - 1;
1629 s->block_index[1]= s->block_wrap[0]*(mb_y*2 + 1);
1630 s->block_index[2]= s->block_wrap[0]*(mb_y*2 + 2) - 1;
1631 s->block_index[3]= s->block_wrap[0]*(mb_y*2 + 2);
1632 s->block_index[4]= s->block_wrap[4]*(mb_y + 1) + s->block_wrap[0]*(s->mb_height*2 + 2);
1633 s->block_index[5]= s->block_wrap[4]*(mb_y + 1 + s->mb_height + 2) + s->block_wrap[0]*(s->mb_height*2 + 2);
1634 for(mb_x=0; mb_x < s->mb_width; mb_x++) {
1635 /*const */int mb_type= s->mb_type[mb_y * s->mb_width + mb_x];
1643 s->block_index[0]+=2;
1644 s->block_index[1]+=2;
1645 s->block_index[2]+=2;
1646 s->block_index[3]+=2;
1647 s->block_index[4]++;
1648 s->block_index[5]++;
1649 if(mb_type & (mb_type-1)){ // more than 1 MB type possible
1651 s->mv_dir = MV_DIR_FORWARD;
1653 copy_context_before_encode(&backup_s, s, -1);
1655 if(mb_type&MB_TYPE_INTER){
1656 int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
1657 s->mv_type = MV_TYPE_16X16;
1659 s->mv[0][0][0] = s->p_mv_table[xy][0];
1660 s->mv[0][0][1] = s->p_mv_table[xy][1];
1661 init_put_bits(&s->pb, bit_buf[1], 3000, NULL, NULL);
1662 s->block= s->inter_block;
1664 encode_mb(s, s->mv[0][0][0], s->mv[0][0][1]);
1665 d= get_bit_count(&s->pb);
1667 flush_put_bits(&s->pb);
1669 copy_context_after_encode(&best_s, s, MB_TYPE_INTER);
1673 if(mb_type&MB_TYPE_INTER4V){
1674 copy_context_before_encode(s, &backup_s, MB_TYPE_INTER4V);
1675 s->mv_type = MV_TYPE_8X8;
1678 s->mv[0][i][0] = s->motion_val[s->block_index[i]][0];
1679 s->mv[0][i][1] = s->motion_val[s->block_index[i]][1];
1681 init_put_bits(&s->pb, bit_buf[2], 3000, NULL, NULL);
1682 s->block= s->inter4v_block;
1685 d= get_bit_count(&s->pb);
1687 flush_put_bits(&s->pb);
1689 copy_context_after_encode(&best_s, s, MB_TYPE_INTER4V);
1693 if(mb_type&MB_TYPE_INTRA){
1694 copy_context_before_encode(s, &backup_s, MB_TYPE_INTRA);
1695 s->mv_type = MV_TYPE_16X16;
1699 init_put_bits(&s->pb, bit_buf[0], 3000, NULL, NULL);
1700 s->block= s->intra_block;
1703 d= get_bit_count(&s->pb);
1705 flush_put_bits(&s->pb);
1707 copy_context_after_encode(&best_s, s, MB_TYPE_INTRA);
1710 /* force cleaning of ac/dc pred stuff if needed ... */
1711 if(s->h263_pred || s->h263_aic)
1712 s->mbintra_table[mb_x + mb_y*s->mb_width]=1;
1714 copy_context_after_encode(s, &best_s, -1);
1715 copy_bits(&pb, bit_buf[best], dmin);
1718 int motion_x, motion_y;
1719 s->mv_type=MV_TYPE_16X16;
1720 // only one MB-Type possible
1721 //FIXME convert to swicth()
1722 if(mb_type&MB_TYPE_INTRA){
1723 s->mv_dir = MV_DIR_FORWARD;
1725 motion_x= s->mv[0][0][0] = 0;
1726 motion_y= s->mv[0][0][1] = 0;
1727 }else if(mb_type&MB_TYPE_INTER){
1728 int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
1729 s->mv_dir = MV_DIR_FORWARD;
1731 motion_x= s->mv[0][0][0] = s->p_mv_table[xy][0];
1732 motion_y= s->mv[0][0][1] = s->p_mv_table[xy][1];
1733 }else if(mb_type&MB_TYPE_DIRECT){
1734 int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
1735 s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD | MV_DIRECT;
1743 }else if(mb_type&MB_TYPE_BIDIR){
1744 int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
1745 s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
1749 s->mv[0][0][0] = s->b_bidir_forw_mv_table[xy][0];
1750 s->mv[0][0][1] = s->b_bidir_forw_mv_table[xy][1];
1751 s->mv[1][0][0] = s->b_bidir_back_mv_table[xy][0];
1752 s->mv[1][0][1] = s->b_bidir_back_mv_table[xy][1];
1753 }else if(mb_type&MB_TYPE_BACKWARD){
1754 int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
1755 s->mv_dir = MV_DIR_BACKWARD;
1757 motion_x= s->mv[1][0][0] = s->b_back_mv_table[xy][0];
1758 motion_y= s->mv[1][0][1] = s->b_back_mv_table[xy][1];
1759 }else if(mb_type&MB_TYPE_FORWARD){
1760 int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
1761 s->mv_dir = MV_DIR_FORWARD;
1763 motion_x= s->mv[0][0][0] = s->b_forw_mv_table[xy][0];
1764 motion_y= s->mv[0][0][1] = s->b_forw_mv_table[xy][1];
1765 // printf(" %d %d ", motion_x, motion_y);
1767 motion_x=motion_y=0; //gcc warning fix
1768 printf("illegal MB type\n");
1770 encode_mb(s, motion_x, motion_y);
1773 MPV_decode_mb(s, s->block);
1777 /* Obtain average GOB size for RTP */
1780 s->mb_line_avgsize = pbBufPtr(&s->pb) - s->ptr_last_mb_line;
1781 else if (!(mb_y % s->gob_index)) {
1782 s->mb_line_avgsize = (s->mb_line_avgsize + pbBufPtr(&s->pb) - s->ptr_last_mb_line) >> 1;
1783 s->ptr_last_mb_line = pbBufPtr(&s->pb);
1785 //fprintf(stderr, "\nMB line: %d\tSize: %u\tAvg. Size: %u", s->mb_y,
1786 // (s->pb.buf_ptr - s->ptr_last_mb_line), s->mb_line_avgsize);
1787 s->first_gob_line = 0;
1792 if (s->h263_msmpeg4 && s->msmpeg4_version<4 && s->pict_type == I_TYPE)
1793 msmpeg4_encode_ext_header(s);
1795 //if (s->gob_number)
1796 // fprintf(stderr,"\nNumber of GOB: %d", s->gob_number);
1798 /* Send the last GOB if RTP */
1800 flush_put_bits(&s->pb);
1801 pdif = pbBufPtr(&s->pb) - s->ptr_lastgob;
1802 /* Call the RTP callback to send the last GOB */
1803 if (s->rtp_callback)
1804 s->rtp_callback(s->ptr_lastgob, pdif, s->gob_number);
1805 s->ptr_lastgob = pbBufPtr(&s->pb);
1806 //fprintf(stderr,"\nGOB: %2d size: %d (last)", s->gob_number, pdif);
1810 static int dct_quantize_c(MpegEncContext *s,
1811 DCTELEM *block, int n,
1814 int i, j, level, last_non_zero, q;
1816 int minLevel, maxLevel;
1818 if(s->avctx!=NULL && s->avctx->codec->id==CODEC_ID_MPEG4){
1822 }else if(s->out_format==FMT_MPEG1){
1826 }else if(s->out_format==FMT_MJPEG){
1831 /* h263 / msmpeg4 */
1838 /* we need this permutation so that we correct the IDCT
1839 permutation. will be moved into DCT code */
1840 block_permute(block);
1849 /* note: block[0] is assumed to be positive */
1850 block[0] = (block[0] + (q >> 1)) / q;
1853 if (s->out_format == FMT_H263) {
1854 qmat = s->q_non_intra_matrix;
1856 qmat = s->q_intra_matrix;
1861 qmat = s->q_non_intra_matrix;
1865 j = zigzag_direct[i];
1867 level = level * qmat[j];
1870 static int count = 0;
1871 int level1, level2, qmat1;
1873 if (qmat == s->q_non_intra_matrix) {
1874 qmat1 = default_non_intra_matrix[j] * s->qscale;
1876 qmat1 = default_intra_matrix[j] * s->qscale;
1878 if (av_fdct != jpeg_fdct_ifast)
1879 val = ((double)block[j] * 8.0) / (double)qmat1;
1881 val = ((double)block[j] * 8.0 * 2048.0) /
1882 ((double)qmat1 * aanscales[j]);
1884 level2 = level / (1 << (QMAT_SHIFT - 3));
1885 if (level1 != level2) {
1886 fprintf(stderr, "%d: quant error qlevel=%d wanted=%d level=%d qmat1=%d qmat=%d wantedf=%0.6f\n",
1887 count, level2, level1, block[j], qmat1, qmat[j],
1894 /* XXX: slight error for the low range. Test should be equivalent to
1895 (level <= -(1 << (QMAT_SHIFT - 3)) || level >= (1 <<
1898 if (((level << (31 - (QMAT_SHIFT - 3))) >> (31 - (QMAT_SHIFT - 3))) !=
1900 level = level / (1 << (QMAT_SHIFT - 3));
1901 /* XXX: currently, this code is not optimal. the range should be:
1907 if (level > maxLevel)
1909 else if (level < minLevel)
1918 return last_non_zero;
1921 static void dct_unquantize_mpeg1_c(MpegEncContext *s,
1922 DCTELEM *block, int n, int qscale)
1924 int i, level, nCoeffs;
1925 const UINT16 *quant_matrix;
1927 if(s->alternate_scan) nCoeffs= 64;
1928 else nCoeffs= s->block_last_index[n]+1;
1932 block[0] = block[0] * s->y_dc_scale;
1934 block[0] = block[0] * s->c_dc_scale;
1935 /* XXX: only mpeg1 */
1936 quant_matrix = s->intra_matrix;
1937 for(i=1;i<nCoeffs;i++) {
1938 int j= zigzag_direct[i];
1943 level = (int)(level * qscale * quant_matrix[j]) >> 3;
1944 level = (level - 1) | 1;
1947 level = (int)(level * qscale * quant_matrix[j]) >> 3;
1948 level = (level - 1) | 1;
1951 if (level < -2048 || level > 2047)
1952 fprintf(stderr, "unquant error %d %d\n", i, level);
1959 quant_matrix = s->non_intra_matrix;
1960 for(;i<nCoeffs;i++) {
1961 int j= zigzag_direct[i];
1966 level = (((level << 1) + 1) * qscale *
1967 ((int) (quant_matrix[j]))) >> 4;
1968 level = (level - 1) | 1;
1971 level = (((level << 1) + 1) * qscale *
1972 ((int) (quant_matrix[j]))) >> 4;
1973 level = (level - 1) | 1;
1976 if (level < -2048 || level > 2047)
1977 fprintf(stderr, "unquant error %d %d\n", i, level);
1985 static void dct_unquantize_mpeg2_c(MpegEncContext *s,
1986 DCTELEM *block, int n, int qscale)
1988 int i, level, nCoeffs;
1989 const UINT16 *quant_matrix;
1991 if(s->alternate_scan) nCoeffs= 64;
1992 else nCoeffs= s->block_last_index[n]+1;
1996 block[0] = block[0] * s->y_dc_scale;
1998 block[0] = block[0] * s->c_dc_scale;
1999 quant_matrix = s->intra_matrix;
2000 for(i=1;i<nCoeffs;i++) {
2001 int j= zigzag_direct[i];
2006 level = (int)(level * qscale * quant_matrix[j]) >> 3;
2009 level = (int)(level * qscale * quant_matrix[j]) >> 3;
2012 if (level < -2048 || level > 2047)
2013 fprintf(stderr, "unquant error %d %d\n", i, level);
2021 quant_matrix = s->non_intra_matrix;
2022 for(;i<nCoeffs;i++) {
2023 int j= zigzag_direct[i];
2028 level = (((level << 1) + 1) * qscale *
2029 ((int) (quant_matrix[j]))) >> 4;
2032 level = (((level << 1) + 1) * qscale *
2033 ((int) (quant_matrix[j]))) >> 4;
2036 if (level < -2048 || level > 2047)
2037 fprintf(stderr, "unquant error %d %d\n", i, level);
2048 static void dct_unquantize_h263_c(MpegEncContext *s,
2049 DCTELEM *block, int n, int qscale)
2051 int i, level, qmul, qadd;
2057 block[0] = block[0] * s->y_dc_scale;
2059 block[0] = block[0] * s->c_dc_scale;
2062 nCoeffs= 64; //does not allways use zigzag table
2065 nCoeffs= zigzag_end[ s->block_last_index[n] ];
2068 qmul = s->qscale << 1;
2069 if (s->h263_aic && s->mb_intra)
2072 qadd = (s->qscale - 1) | 1;
2074 for(;i<nCoeffs;i++) {
2078 level = level * qmul - qadd;
2080 level = level * qmul + qadd;
2083 if (level < -2048 || level > 2047)
2084 fprintf(stderr, "unquant error %d %d\n", i, level);
2093 static void rate_control_init(MpegEncContext *s)
2097 //initial values, they dont really matter as they will be totally different within a few frames
2098 s->i_pred.coeff= s->p_pred.coeff= 7.0;
2099 s->i_pred.count= s->p_pred.count= 1.0;
2101 s->i_pred.decay= s->p_pred.decay= 0.4;
2103 // use more bits at the beginning, otherwise high motion at the begin will look like shit
2107 s->short_term_qsum=0.001;
2108 s->short_term_qcount=0.001;
2111 static double predict(Predictor *p, double q, double var)
2113 return p->coeff*var / (q*p->count);
2116 static void update_predictor(Predictor *p, double q, double var, double size)
2118 double new_coeff= size*q / (var + 1);
2119 if(var<1000) return;
2121 int pred= predict(p, q, var);
2122 int error= abs(pred-size);
2123 static double sum=0;
2125 if(count>5) sum+=error;
2127 if(256*256*256*64%count==0){
2128 printf("%d %f %f\n", count, sum/count, p->coeff);
2131 p->count*= p->decay;
2132 p->coeff*= p->decay;
2134 p->coeff+= new_coeff;
2137 static int rate_estimate_qscale(MpegEncContext *s)
2144 float br_compensation;
2146 double short_term_q;
2148 int last_qscale= s->qscale;
2153 fps= (double)s->frame_rate / FRAME_RATE_BASE;
2154 wanted_bits= s->bit_rate*(double)s->picture_number/fps;
2157 if(s->picture_number>2){
2158 /* update predictors */
2159 if(s->last_pict_type == I_TYPE){
2162 //printf("%d %d %d %f\n", s->qscale, s->last_mc_mb_var, s->frame_bits, s->p_pred.coeff);
2163 update_predictor(&s->p_pred, s->qscale, s->last_mc_mb_var, s->frame_bits);
2167 if(s->pict_type == I_TYPE){
2169 rate_q= s->qsum/s->qcount;
2172 int diff, best_diff=1000000000;
2173 for(i=1; i<=31; i++){
2174 diff= predict(&s->p_pred, i, s->mc_mb_var) - (double)s->bit_rate/fps;
2175 if(diff<0) diff= -diff;
2183 s->short_term_qsum*=s->qblur;
2184 s->short_term_qcount*=s->qblur;
2186 s->short_term_qsum+= rate_q;
2187 s->short_term_qcount++;
2188 short_term_q= s->short_term_qsum/s->short_term_qcount;
2190 long_term_q= s->qsum/s->qcount*(s->total_bits+1)/(wanted_bits+1); //+1 to avoid nan & 0
2192 // q= (long_term_q - short_term_q)*s->qcompress + short_term_q;
2193 q= 1/((1/long_term_q - 1/short_term_q)*s->qcompress + 1/short_term_q);
2195 diff= s->total_bits - wanted_bits;
2196 br_compensation= (s->bit_rate_tolerance - diff)/s->bit_rate_tolerance;
2197 if(br_compensation<=0.0) br_compensation=0.001;
2200 qscale= (int)(q + 0.5);
2201 if (qscale<qmin) qscale=qmin;
2202 else if(qscale>qmax) qscale=qmax;
2204 if (qscale<last_qscale-s->max_qdiff) qscale=last_qscale-s->max_qdiff;
2205 else if(qscale>last_qscale+s->max_qdiff) qscale=last_qscale+s->max_qdiff;
2210 s->last_pict_type= s->pict_type;
2211 //printf("q:%d diff:%d comp:%f rate_q:%d st_q:%f fvar:%d last_size:%d\n", qscale, (int)diff, br_compensation,
2212 // rate_q, short_term_q, s->mc_mb_var, s->frame_bits);
2213 //printf("%d %d\n", s->bit_rate, (int)fps);
2217 AVCodec mpeg1video_encoder = {
2220 CODEC_ID_MPEG1VIDEO,
2221 sizeof(MpegEncContext),
2227 AVCodec h263_encoder = {
2231 sizeof(MpegEncContext),
2237 AVCodec h263p_encoder = {
2241 sizeof(MpegEncContext),
2247 AVCodec rv10_encoder = {
2251 sizeof(MpegEncContext),
2257 AVCodec mjpeg_encoder = {
2261 sizeof(MpegEncContext),
2267 AVCodec mpeg4_encoder = {
2271 sizeof(MpegEncContext),
2277 AVCodec msmpeg4v1_encoder = {
2281 sizeof(MpegEncContext),
2287 AVCodec msmpeg4v2_encoder = {
2291 sizeof(MpegEncContext),
2297 AVCodec msmpeg4v3_encoder = {
2301 sizeof(MpegEncContext),