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.
25 #include "mpegvideo.h"
28 #include "fastmemcpy.h"
31 static void encode_picture(MpegEncContext *s, int picture_number);
32 static void rate_control_init(MpegEncContext *s);
33 static int rate_estimate_qscale(MpegEncContext *s);
34 static void dct_unquantize_mpeg1_c(MpegEncContext *s,
35 DCTELEM *block, int n, int qscale);
36 static void dct_unquantize_h263_c(MpegEncContext *s,
37 DCTELEM *block, int n, int qscale);
38 static void draw_edges_c(UINT8 *buf, int wrap, int width, int height, int w);
39 static int dct_quantize_c(MpegEncContext *s, DCTELEM *block, int n, int qscale);
41 int (*dct_quantize)(MpegEncContext *s, DCTELEM *block, int n, int qscale)= dct_quantize_c;
42 void (*draw_edges)(UINT8 *buf, int wrap, int width, int height, int w)= draw_edges_c;
46 /* enable all paranoid tests for rounding, overflows, etc... */
51 /* for jpeg fast DCT */
54 static const unsigned short aanscales[64] = {
55 /* precomputed values scaled up by 14 bits */
56 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
57 22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270,
58 21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906,
59 19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315,
60 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
61 12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552,
62 8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446,
63 4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247
66 static UINT8 h263_chroma_roundtab[16] = {
67 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2,
70 static UINT16 default_mv_penalty[MAX_FCODE+1][MAX_MV*2+1];
71 static UINT8 default_fcode_tab[MAX_MV*2+1];
73 /* default motion estimation */
74 int motion_estimation_method = ME_LOG;
76 extern UINT8 zigzag_end[64];
78 static void convert_matrix(int *qmat, UINT16 *qmat16, const UINT16 *quant_matrix, int qscale)
82 if (av_fdct == jpeg_fdct_ifast) {
84 /* 16 <= qscale * quant_matrix[i] <= 7905 */
85 /* 19952 <= aanscales[i] * qscale * quant_matrix[i] <= 249205026 */
86 /* (1<<36)/19952 >= (1<<36)/(aanscales[i] * qscale * quant_matrix[i]) >= (1<<36)/249205026 */
87 /* 3444240 >= (1<<36)/(aanscales[i] * qscale * quant_matrix[i]) >= 275 */
89 qmat[block_permute_op(i)] = (int)((UINT64_C(1) << (QMAT_SHIFT + 11)) /
90 (aanscales[i] * qscale * quant_matrix[block_permute_op(i)]));
94 /* We can safely suppose that 16 <= quant_matrix[i] <= 255
95 So 16 <= qscale * quant_matrix[i] <= 7905
96 so (1<<19) / 16 >= (1<<19) / (qscale * quant_matrix[i]) >= (1<<19) / 7905
97 so 32768 >= (1<<19) / (qscale * quant_matrix[i]) >= 67
99 qmat[i] = (1 << QMAT_SHIFT_MMX) / (qscale * quant_matrix[i]);
100 qmat16[i] = (1 << QMAT_SHIFT_MMX) / (qscale * quant_matrix[block_permute_op(i)]);
105 /* init common structure for both encoder and decoder */
106 int MPV_common_init(MpegEncContext *s)
111 if (s->out_format == FMT_H263)
112 s->dct_unquantize = dct_unquantize_h263_c;
114 s->dct_unquantize = dct_unquantize_mpeg1_c;
117 MPV_common_init_mmx(s);
119 s->mb_width = (s->width + 15) / 16;
120 s->mb_height = (s->height + 15) / 16;
121 s->mb_num = s->mb_width * s->mb_height;
122 s->linesize = s->mb_width * 16 + 2 * EDGE_WIDTH;
125 int w, h, shift, pict_start;
128 h = s->mb_height * 16 + 2 * EDGE_WIDTH;
129 shift = (i == 0) ? 0 : 1;
130 c_size = (w >> shift) * (h >> shift);
131 pict_start = (w >> shift) * (EDGE_WIDTH >> shift) + (EDGE_WIDTH >> shift);
133 pict = av_mallocz(c_size);
136 s->last_picture_base[i] = pict;
137 s->last_picture[i] = pict + pict_start;
139 pict = av_mallocz(c_size);
142 s->next_picture_base[i] = pict;
143 s->next_picture[i] = pict + pict_start;
145 if (s->has_b_frames) {
146 pict = av_mallocz(c_size);
149 s->aux_picture_base[i] = pict;
150 s->aux_picture[i] = pict + pict_start;
155 /* Allocate MB type table */
156 s->mb_type = av_mallocz(s->mb_num * sizeof(char));
157 if (s->mb_type == NULL) {
162 s->mb_var = av_mallocz(s->mb_num * sizeof(INT16));
163 if (s->mb_var == NULL) {
167 /* Allocate MV table */
168 /* By now we just have one MV per MB */
169 s->mv_table[0] = av_mallocz(s->mb_num * sizeof(INT16));
170 s->mv_table[1] = av_mallocz(s->mb_num * sizeof(INT16));
171 if (s->mv_table[1] == NULL || s->mv_table[0] == NULL) {
177 if (s->out_format == FMT_H263 || s->encoding) {
180 size = (2 * s->mb_width + 2) * (2 * s->mb_height + 2);
181 s->motion_val = malloc(size * 2 * sizeof(INT16));
182 if (s->motion_val == NULL)
184 memset(s->motion_val, 0, size * 2 * sizeof(INT16));
187 if (s->h263_pred || s->h263_plus) {
188 int y_size, c_size, i, size;
192 y_size = (2 * s->mb_width + 2) * (2 * s->mb_height + 2);
193 c_size = (s->mb_width + 2) * (s->mb_height + 2);
194 size = y_size + 2 * c_size;
195 s->dc_val[0] = malloc(size * sizeof(INT16));
196 if (s->dc_val[0] == NULL)
198 s->dc_val[1] = s->dc_val[0] + y_size;
199 s->dc_val[2] = s->dc_val[1] + c_size;
201 s->dc_val[0][i] = 1024;
204 s->ac_val[0] = av_mallocz(size * sizeof(INT16) * 16);
205 if (s->ac_val[0] == NULL)
207 s->ac_val[1] = s->ac_val[0] + y_size;
208 s->ac_val[2] = s->ac_val[1] + c_size;
211 s->coded_block = av_mallocz(y_size);
215 /* which mb is a intra block */
216 s->mbintra_table = av_mallocz(s->mb_num);
217 if (!s->mbintra_table)
219 memset(s->mbintra_table, 1, s->mb_num);
221 /* default structure is frame */
222 s->picture_structure = PICT_FRAME;
224 /* init macroblock skip table */
226 s->mbskip_table = av_mallocz(s->mb_num);
227 if (!s->mbskip_table)
231 s->block= s->intra_block;
233 s->context_initialized = 1;
240 /* init common structure for both encoder and decoder */
241 void MPV_common_end(MpegEncContext *s)
250 free(s->mv_table[0]);
252 free(s->mv_table[1]);
260 free(s->coded_block);
261 if (s->mbintra_table)
262 free(s->mbintra_table);
265 free(s->mbskip_table);
267 if (s->last_picture_base[i])
268 free(s->last_picture_base[i]);
269 if (s->next_picture_base[i])
270 free(s->next_picture_base[i]);
272 free(s->aux_picture_base[i]);
274 s->context_initialized = 0;
277 /* init video encoder */
278 int MPV_encode_init(AVCodecContext *avctx)
280 MpegEncContext *s = avctx->priv_data;
283 s->bit_rate = avctx->bit_rate;
284 s->bit_rate_tolerance = avctx->bit_rate_tolerance;
285 s->frame_rate = avctx->frame_rate;
286 s->width = avctx->width;
287 s->height = avctx->height;
288 s->gop_size = avctx->gop_size;
289 s->rtp_mode = avctx->rtp_mode;
290 s->rtp_payload_size = avctx->rtp_payload_size;
291 if (avctx->rtp_callback)
292 s->rtp_callback = avctx->rtp_callback;
293 s->qmin= avctx->qmin;
294 s->qmax= avctx->qmax;
295 s->max_qdiff= avctx->max_qdiff;
296 s->qcompress= avctx->qcompress;
297 s->qblur= avctx->qblur;
299 s->aspect_ratio_info= avctx->aspect_ratio_info;
300 s->flags= avctx->flags;
302 if (s->gop_size <= 1) {
308 s->full_search = motion_estimation_method;
310 s->fixed_qscale = (avctx->flags & CODEC_FLAG_QSCALE);
312 switch(avctx->codec->id) {
313 case CODEC_ID_MPEG1VIDEO:
314 s->out_format = FMT_MPEG1;
317 s->out_format = FMT_MJPEG;
318 s->intra_only = 1; /* force intra only for jpeg */
319 s->mjpeg_write_tables = 1; /* write all tables */
320 s->mjpeg_vsample[0] = 2; /* set up default sampling factors */
321 s->mjpeg_vsample[1] = 1; /* the only currently supported values */
322 s->mjpeg_vsample[2] = 1;
323 s->mjpeg_hsample[0] = 2;
324 s->mjpeg_hsample[1] = 1;
325 s->mjpeg_hsample[2] = 1;
326 if (mjpeg_init(s) < 0)
330 if (h263_get_picture_format(s->width, s->height) == 7) {
331 printf("Input picture size isn't suitable for h263 codec! try h263+\n");
334 s->out_format = FMT_H263;
337 s->out_format = FMT_H263;
339 s->rtp_payload_size = 1200;
341 s->unrestricted_mv = 1;
343 /* These are just to be sure */
348 s->out_format = FMT_H263;
352 s->out_format = FMT_H263;
354 s->unrestricted_mv = 1;
356 case CODEC_ID_MSMPEG4:
357 s->out_format = FMT_H263;
360 s->unrestricted_mv = 1;
366 { /* set up some save defaults, some codecs might override them later */
371 memset(default_mv_penalty, 0, sizeof(UINT16)*(MAX_FCODE+1)*(2*MAX_MV+1));
372 memset(default_fcode_tab , 0, sizeof(UINT8)*(2*MAX_MV+1));
374 for(i=-16; i<16; i++){
375 default_fcode_tab[i + MAX_MV]= 1;
379 s->mv_penalty= default_mv_penalty;
380 s->fcode_tab= default_fcode_tab;
382 if (s->out_format == FMT_H263)
384 else if (s->out_format == FMT_MPEG1)
385 mpeg1_encode_init(s);
387 /* dont use mv_penalty table for crap MV as it would be confused */
388 if(s->full_search<4) s->mv_penalty= default_mv_penalty;
393 if (MPV_common_init(s) < 0)
396 /* init default q matrix */
398 s->intra_matrix[i] = default_intra_matrix[i];
399 s->non_intra_matrix[i] = default_non_intra_matrix[i];
402 /* rate control init */
403 rate_control_init(s);
405 s->picture_number = 0;
406 s->picture_in_gop_number = 0;
407 s->fake_picture_number = 0;
408 /* motion detector init */
414 int MPV_encode_end(AVCodecContext *avctx)
416 MpegEncContext *s = avctx->priv_data;
422 if (s->out_format == FMT_MJPEG)
428 /* draw the edges of width 'w' of an image of size width, height */
429 static void draw_edges_c(UINT8 *buf, int wrap, int width, int height, int w)
431 UINT8 *ptr, *last_line;
434 last_line = buf + (height - 1) * wrap;
437 memcpy(buf - (i + 1) * wrap, buf, width);
438 memcpy(last_line + (i + 1) * wrap, last_line, width);
442 for(i=0;i<height;i++) {
443 memset(ptr - w, ptr[0], w);
444 memset(ptr + width, ptr[width-1], w);
449 memset(buf - (i + 1) * wrap - w, buf[0], w); /* top left */
450 memset(buf - (i + 1) * wrap + width, buf[width-1], w); /* top right */
451 memset(last_line + (i + 1) * wrap - w, last_line[0], w); /* top left */
452 memset(last_line + (i + 1) * wrap + width, last_line[width-1], w); /* top right */
456 /* generic function for encode/decode called before a frame is coded/decoded */
457 void MPV_frame_start(MpegEncContext *s)
463 if (s->pict_type == B_TYPE) {
465 s->current_picture[i] = s->aux_picture[i];
468 s->last_non_b_pict_type= s->pict_type;
470 /* swap next and last */
471 tmp = s->last_picture[i];
472 s->last_picture[i] = s->next_picture[i];
473 s->next_picture[i] = tmp;
474 s->current_picture[i] = tmp;
479 /* generic function for encode/decode called after a frame has been coded/decoded */
480 void MPV_frame_end(MpegEncContext *s)
482 /* draw edge for correct motion prediction if outside */
483 if (s->pict_type != B_TYPE && !s->intra_only) {
484 if(s->avctx==NULL || s->avctx->codec->id!=CODEC_ID_MPEG4 || s->divx_version==500){
485 draw_edges(s->current_picture[0], s->linesize, s->mb_width*16, s->mb_height*16, EDGE_WIDTH);
486 draw_edges(s->current_picture[1], s->linesize/2, s->mb_width*8, s->mb_height*8, EDGE_WIDTH/2);
487 draw_edges(s->current_picture[2], s->linesize/2, s->mb_width*8, s->mb_height*8, EDGE_WIDTH/2);
489 /* mpeg4? / opendivx / xvid */
490 draw_edges(s->current_picture[0], s->linesize, s->width, s->height, EDGE_WIDTH);
491 draw_edges(s->current_picture[1], s->linesize/2, s->width/2, s->height/2, EDGE_WIDTH/2);
492 draw_edges(s->current_picture[2], s->linesize/2, s->width/2, s->height/2, EDGE_WIDTH/2);
498 int MPV_encode_picture(AVCodecContext *avctx,
499 unsigned char *buf, int buf_size, void *data)
501 MpegEncContext *s = avctx->priv_data;
502 AVPicture *pict = data;
506 s->qscale = avctx->quality;
508 init_put_bits(&s->pb, buf, buf_size, NULL, NULL);
510 if (!s->intra_only) {
511 /* first picture of GOP is intra */
512 if (s->picture_in_gop_number % s->gop_size==0){
513 s->picture_in_gop_number=0;
514 s->pict_type = I_TYPE;
516 s->pict_type = P_TYPE;
518 s->pict_type = I_TYPE;
524 UINT8 *src = pict->data[i];
525 UINT8 *dest = s->current_picture[i];
526 int src_wrap = pict->linesize[i];
527 int dest_wrap = s->linesize;
537 if(dest_wrap==src_wrap){
538 s->new_picture[i] = pict->data[i];
541 memcpy(dest, src, w);
545 s->new_picture[i] = s->current_picture[i];
549 encode_picture(s, s->picture_number);
550 avctx->key_frame = (s->pict_type == I_TYPE);
551 avctx->header_bits = s->header_bits;
552 avctx->mv_bits = s->mv_bits;
553 avctx->misc_bits = s->misc_bits;
554 avctx->i_tex_bits = s->i_tex_bits;
555 avctx->p_tex_bits = s->p_tex_bits;
556 avctx->i_count = s->i_count;
557 avctx->p_count = s->p_count;
558 avctx->skip_count = s->skip_count;
562 s->picture_in_gop_number++;
564 if (s->out_format == FMT_MJPEG)
565 mjpeg_picture_trailer(s);
567 flush_put_bits(&s->pb);
568 s->last_frame_bits= s->frame_bits;
569 s->frame_bits = (pbBufPtr(&s->pb) - s->pb.buf) * 8;
570 s->total_bits += s->frame_bits;
571 avctx->frame_bits = s->frame_bits;
572 //printf("fcode: %d, type: %d, head: %d, mv: %d, misc: %d, frame: %d, itex: %d, ptex: %d\n",
573 //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);
575 avctx->quality = s->qscale;
576 if (avctx->get_psnr) {
577 /* At this point pict->data should have the original frame */
578 /* an s->current_picture should have the coded/decoded frame */
579 get_psnr(pict->data, s->current_picture,
580 pict->linesize, s->linesize, avctx);
582 return pbBufPtr(&s->pb) - s->pb.buf;
585 static inline int clip(int a, int amin, int amax)
595 static inline void gmc1_motion(MpegEncContext *s,
596 UINT8 *dest_y, UINT8 *dest_cb, UINT8 *dest_cr,
598 UINT8 **ref_picture, int src_offset,
602 int dxy, offset, mx, my, src_x, src_y, height, linesize;
603 int motion_x, motion_y;
605 if(s->real_sprite_warping_points>1) printf("more than 1 warp point isnt supported\n");
606 motion_x= s->sprite_offset[0][0];
607 motion_y= s->sprite_offset[0][1];
608 src_x = s->mb_x * 16 + (motion_x >> (s->sprite_warping_accuracy+1));
609 src_y = s->mb_y * 16 + (motion_y >> (s->sprite_warping_accuracy+1));
610 motion_x<<=(3-s->sprite_warping_accuracy);
611 motion_y<<=(3-s->sprite_warping_accuracy);
612 src_x = clip(src_x, -16, s->width);
613 if (src_x == s->width)
615 src_y = clip(src_y, -16, s->height);
616 if (src_y == s->height)
619 linesize = s->linesize;
620 ptr = ref_picture[0] + (src_y * linesize) + src_x + src_offset;
623 gmc1(dest_y , ptr , linesize, h, motion_x&15, motion_y&15, s->no_rounding);
624 gmc1(dest_y+8, ptr+8, linesize, h, motion_x&15, motion_y&15, s->no_rounding);
626 motion_x= s->sprite_offset[1][0];
627 motion_y= s->sprite_offset[1][1];
628 src_x = s->mb_x * 8 + (motion_x >> (s->sprite_warping_accuracy+1));
629 src_y = s->mb_y * 8 + (motion_y >> (s->sprite_warping_accuracy+1));
630 motion_x<<=(3-s->sprite_warping_accuracy);
631 motion_y<<=(3-s->sprite_warping_accuracy);
632 src_x = clip(src_x, -8, s->width>>1);
633 if (src_x == s->width>>1)
635 src_y = clip(src_y, -8, s->height>>1);
636 if (src_y == s->height>>1)
639 offset = (src_y * linesize>>1) + src_x + (src_offset>>1);
640 ptr = ref_picture[1] + offset;
641 gmc1(dest_cb + (dest_offset>>1), ptr, linesize>>1, h>>1, motion_x&15, motion_y&15, s->no_rounding);
642 ptr = ref_picture[2] + offset;
643 gmc1(dest_cr + (dest_offset>>1), ptr, linesize>>1, h>>1, motion_x&15, motion_y&15, s->no_rounding);
648 /* apply one mpeg motion vector to the three components */
649 static inline void mpeg_motion(MpegEncContext *s,
650 UINT8 *dest_y, UINT8 *dest_cb, UINT8 *dest_cr,
652 UINT8 **ref_picture, int src_offset,
653 int field_based, op_pixels_func *pix_op,
654 int motion_x, int motion_y, int h)
657 int dxy, offset, mx, my, src_x, src_y, height, linesize;
658 if(s->quarter_sample)
663 dxy = ((motion_y & 1) << 1) | (motion_x & 1);
664 src_x = s->mb_x * 16 + (motion_x >> 1);
665 src_y = s->mb_y * (16 >> field_based) + (motion_y >> 1);
667 /* WARNING: do no forget half pels */
668 height = s->height >> field_based;
669 src_x = clip(src_x, -16, s->width);
670 if (src_x == s->width)
672 src_y = clip(src_y, -16, height);
675 linesize = s->linesize << field_based;
676 ptr = ref_picture[0] + (src_y * linesize) + (src_x) + src_offset;
677 dest_y += dest_offset;
678 pix_op[dxy](dest_y, ptr, linesize, h);
679 pix_op[dxy](dest_y + 8, ptr + 8, linesize, h);
681 if (s->out_format == FMT_H263) {
683 if ((motion_x & 3) != 0)
685 if ((motion_y & 3) != 0)
692 dxy = ((my & 1) << 1) | (mx & 1);
697 src_x = s->mb_x * 8 + mx;
698 src_y = s->mb_y * (8 >> field_based) + my;
699 src_x = clip(src_x, -8, s->width >> 1);
700 if (src_x == (s->width >> 1))
702 src_y = clip(src_y, -8, height >> 1);
703 if (src_y == (height >> 1))
706 offset = (src_y * (linesize >> 1)) + src_x + (src_offset >> 1);
707 ptr = ref_picture[1] + offset;
708 pix_op[dxy](dest_cb + (dest_offset >> 1), ptr, linesize >> 1, h >> 1);
709 ptr = ref_picture[2] + offset;
710 pix_op[dxy](dest_cr + (dest_offset >> 1), ptr, linesize >> 1, h >> 1);
713 static inline void qpel_motion(MpegEncContext *s,
714 UINT8 *dest_y, UINT8 *dest_cb, UINT8 *dest_cr,
716 UINT8 **ref_picture, int src_offset,
717 int field_based, op_pixels_func *pix_op,
718 qpel_mc_func *qpix_op,
719 int motion_x, int motion_y, int h)
722 int dxy, offset, mx, my, src_x, src_y, height, linesize;
724 dxy = ((motion_y & 3) << 2) | (motion_x & 3);
725 src_x = s->mb_x * 16 + (motion_x >> 2);
726 src_y = s->mb_y * (16 >> field_based) + (motion_y >> 2);
728 height = s->height >> field_based;
729 src_x = clip(src_x, -16, s->width);
730 if (src_x == s->width)
732 src_y = clip(src_y, -16, height);
735 linesize = s->linesize << field_based;
736 ptr = ref_picture[0] + (src_y * linesize) + src_x + src_offset;
737 dest_y += dest_offset;
738 //printf("%d %d %d\n", src_x, src_y, dxy);
739 qpix_op[dxy](dest_y , ptr , linesize, linesize, motion_x&3, motion_y&3);
740 qpix_op[dxy](dest_y + 8, ptr + 8, linesize, linesize, motion_x&3, motion_y&3);
741 qpix_op[dxy](dest_y + linesize*8 , ptr + linesize*8 , linesize, linesize, motion_x&3, motion_y&3);
742 qpix_op[dxy](dest_y + linesize*8 + 8, ptr + linesize*8 + 8, linesize, linesize, motion_x&3, motion_y&3);
744 mx= (motion_x>>1) | (motion_x&1);
745 my= (motion_y>>1) | (motion_y&1);
755 src_x = s->mb_x * 8 + mx;
756 src_y = s->mb_y * (8 >> field_based) + my;
757 src_x = clip(src_x, -8, s->width >> 1);
758 if (src_x == (s->width >> 1))
760 src_y = clip(src_y, -8, height >> 1);
761 if (src_y == (height >> 1))
764 offset = (src_y * (linesize >> 1)) + src_x + (src_offset >> 1);
765 ptr = ref_picture[1] + offset;
766 pix_op[dxy](dest_cb + (dest_offset >> 1), ptr, linesize >> 1, h >> 1);
767 ptr = ref_picture[2] + offset;
768 pix_op[dxy](dest_cr + (dest_offset >> 1), ptr, linesize >> 1, h >> 1);
772 static inline void MPV_motion(MpegEncContext *s,
773 UINT8 *dest_y, UINT8 *dest_cb, UINT8 *dest_cr,
774 int dir, UINT8 **ref_picture,
775 op_pixels_func *pix_op, qpel_mc_func *qpix_op)
777 int dxy, offset, mx, my, src_x, src_y, motion_x, motion_y;
788 mpeg_motion(s, dest_y, dest_cb, dest_cr, 0,
791 s->sprite_offset[0][0]>>3,
792 s->sprite_offset[0][1]>>3,
795 gmc1_motion(s, dest_y, dest_cb, dest_cr, 0,
799 }else if(s->quarter_sample && dir==0){ //FIXME
800 qpel_motion(s, dest_y, dest_cb, dest_cr, 0,
803 s->mv[dir][0][0], s->mv[dir][0][1], 16);
805 mpeg_motion(s, dest_y, dest_cb, dest_cr, 0,
808 s->mv[dir][0][0], s->mv[dir][0][1], 16);
813 motion_x = s->mv[dir][i][0];
814 motion_y = s->mv[dir][i][1];
816 dxy = ((motion_y & 1) << 1) | (motion_x & 1);
817 src_x = mb_x * 16 + (motion_x >> 1) + (i & 1) * 8;
818 src_y = mb_y * 16 + (motion_y >> 1) + ((i >> 1) & 1) * 8;
820 /* WARNING: do no forget half pels */
821 src_x = clip(src_x, -16, s->width);
822 if (src_x == s->width)
824 src_y = clip(src_y, -16, s->height);
825 if (src_y == s->height)
828 ptr = ref_picture[0] + (src_y * s->linesize) + (src_x);
829 dest = dest_y + ((i & 1) * 8) + (i >> 1) * 8 * s->linesize;
830 pix_op[dxy](dest, ptr, s->linesize, 8);
832 /* In case of 8X8, we construct a single chroma motion vector
833 with a special rounding */
837 mx += s->mv[dir][i][0];
838 my += s->mv[dir][i][1];
841 mx = (h263_chroma_roundtab[mx & 0xf] + ((mx >> 3) & ~1));
844 mx = -(h263_chroma_roundtab[mx & 0xf] + ((mx >> 3) & ~1));
847 my = (h263_chroma_roundtab[my & 0xf] + ((my >> 3) & ~1));
850 my = -(h263_chroma_roundtab[my & 0xf] + ((my >> 3) & ~1));
852 dxy = ((my & 1) << 1) | (mx & 1);
856 src_x = mb_x * 8 + mx;
857 src_y = mb_y * 8 + my;
858 src_x = clip(src_x, -8, s->width/2);
859 if (src_x == s->width/2)
861 src_y = clip(src_y, -8, s->height/2);
862 if (src_y == s->height/2)
865 offset = (src_y * (s->linesize >> 1)) + src_x;
866 ptr = ref_picture[1] + offset;
867 pix_op[dxy](dest_cb, ptr, s->linesize >> 1, 8);
868 ptr = ref_picture[2] + offset;
869 pix_op[dxy](dest_cr, ptr, s->linesize >> 1, 8);
872 if (s->picture_structure == PICT_FRAME) {
874 mpeg_motion(s, dest_y, dest_cb, dest_cr, 0,
875 ref_picture, s->field_select[dir][0] ? s->linesize : 0,
877 s->mv[dir][0][0], s->mv[dir][0][1], 8);
879 mpeg_motion(s, dest_y, dest_cb, dest_cr, s->linesize,
880 ref_picture, s->field_select[dir][1] ? s->linesize : 0,
882 s->mv[dir][1][0], s->mv[dir][1][1], 8);
892 /* put block[] to dest[] */
893 static inline void put_dct(MpegEncContext *s,
894 DCTELEM *block, int i, UINT8 *dest, int line_size)
897 s->dct_unquantize(s, block, i, s->qscale);
899 put_pixels_clamped(block, dest, line_size);
902 /* add block[] to dest[] */
903 static inline void add_dct(MpegEncContext *s,
904 DCTELEM *block, int i, UINT8 *dest, int line_size)
906 if (s->block_last_index[i] >= 0) {
908 if(s->encoding || (!s->h263_msmpeg4))
909 s->dct_unquantize(s, block, i, s->qscale);
911 add_pixels_clamped(block, dest, line_size);
915 /* generic function called after a macroblock has been parsed by the
916 decoder or after it has been encoded by the encoder.
918 Important variables used:
919 s->mb_intra : true if intra macroblock
920 s->mv_dir : motion vector direction
921 s->mv_type : motion vector type
922 s->mv : motion vector
923 s->interlaced_dct : true if interlaced dct used (mpeg2)
925 void MPV_decode_mb(MpegEncContext *s, DCTELEM block[6][64])
928 int dct_linesize, dct_offset;
929 op_pixels_func *op_pix;
930 qpel_mc_func *op_qpix;
935 #ifdef FF_POSTPROCESS
936 quant_store[mb_y][mb_x]=s->qscale;
937 //printf("[%02d][%02d] %d\n",mb_x,mb_y,s->qscale);
940 /* update DC predictors for P macroblocks */
942 if (s->h263_pred || s->h263_aic) {
943 if(s->mbintra_table[mb_x + mb_y*s->mb_width])
946 s->mbintra_table[mb_x + mb_y*s->mb_width]=0;
947 wrap = 2 * s->mb_width + 2;
948 xy = 2 * mb_x + 1 + (2 * mb_y + 1) * wrap;
951 s->dc_val[0][xy] = v;
952 s->dc_val[0][xy + 1] = v;
953 s->dc_val[0][xy + wrap] = v;
954 s->dc_val[0][xy + 1 + wrap] = v;
956 memset(s->ac_val[0][xy], 0, 16 * sizeof(INT16));
957 memset(s->ac_val[0][xy + 1], 0, 16 * sizeof(INT16));
958 memset(s->ac_val[0][xy + wrap], 0, 16 * sizeof(INT16));
959 memset(s->ac_val[0][xy + 1 + wrap], 0, 16 * sizeof(INT16));
960 if (s->h263_msmpeg4) {
961 s->coded_block[xy] = 0;
962 s->coded_block[xy + 1] = 0;
963 s->coded_block[xy + wrap] = 0;
964 s->coded_block[xy + 1 + wrap] = 0;
967 wrap = s->mb_width + 2;
968 xy = mb_x + 1 + (mb_y + 1) * wrap;
969 s->dc_val[1][xy] = v;
970 s->dc_val[2][xy] = v;
972 memset(s->ac_val[1][xy], 0, 16 * sizeof(INT16));
973 memset(s->ac_val[2][xy], 0, 16 * sizeof(INT16));
976 s->last_dc[0] = 128 << s->intra_dc_precision;
977 s->last_dc[1] = 128 << s->intra_dc_precision;
978 s->last_dc[2] = 128 << s->intra_dc_precision;
981 else if (s->h263_pred || s->h263_aic)
982 s->mbintra_table[mb_x + mb_y*s->mb_width]=1;
984 /* update motion predictor, not for B-frames as they need the motion_val from the last P/S-Frame */
985 if (s->out_format == FMT_H263) {
986 if(s->pict_type!=B_TYPE){
987 int xy, wrap, motion_x, motion_y;
989 wrap = 2 * s->mb_width + 2;
990 xy = 2 * mb_x + 1 + (2 * mb_y + 1) * wrap;
995 } else if (s->mv_type == MV_TYPE_16X16) {
996 motion_x = s->mv[0][0][0];
997 motion_y = s->mv[0][0][1];
999 /* no update if 8X8 because it has been done during parsing */
1000 s->motion_val[xy][0] = motion_x;
1001 s->motion_val[xy][1] = motion_y;
1002 s->motion_val[xy + 1][0] = motion_x;
1003 s->motion_val[xy + 1][1] = motion_y;
1004 s->motion_val[xy + wrap][0] = motion_x;
1005 s->motion_val[xy + wrap][1] = motion_y;
1006 s->motion_val[xy + 1 + wrap][0] = motion_x;
1007 s->motion_val[xy + 1 + wrap][1] = motion_y;
1012 if (!s->intra_only) {
1013 UINT8 *dest_y, *dest_cb, *dest_cr;
1016 /* avoid copy if macroblock skipped in last frame too */
1017 if (!s->encoding && s->pict_type != B_TYPE) {
1018 mbskip_ptr = &s->mbskip_table[s->mb_y * s->mb_width + s->mb_x];
1021 /* if previous was skipped too, then nothing to do ! */
1022 if (*mbskip_ptr != 0)
1024 *mbskip_ptr = 1; /* indicate that this time we skiped it */
1026 *mbskip_ptr = 0; /* not skipped */
1030 dest_y = s->current_picture[0] + (mb_y * 16 * s->linesize) + mb_x * 16;
1031 dest_cb = s->current_picture[1] + (mb_y * 8 * (s->linesize >> 1)) + mb_x * 8;
1032 dest_cr = s->current_picture[2] + (mb_y * 8 * (s->linesize >> 1)) + mb_x * 8;
1034 if (s->interlaced_dct) {
1035 dct_linesize = s->linesize * 2;
1036 dct_offset = s->linesize;
1038 dct_linesize = s->linesize;
1039 dct_offset = s->linesize * 8;
1043 /* motion handling */
1044 if (!s->no_rounding){
1045 op_pix = put_pixels_tab;
1046 op_qpix= qpel_mc_rnd_tab;
1048 op_pix = put_no_rnd_pixels_tab;
1049 op_qpix= qpel_mc_no_rnd_tab;
1052 if (s->mv_dir & MV_DIR_FORWARD) {
1053 MPV_motion(s, dest_y, dest_cb, dest_cr, 0, s->last_picture, op_pix, op_qpix);
1054 if (!s->no_rounding)
1055 op_pix = avg_pixels_tab;
1057 op_pix = avg_no_rnd_pixels_tab;
1059 if (s->mv_dir & MV_DIR_BACKWARD) {
1060 MPV_motion(s, dest_y, dest_cb, dest_cr, 1, s->next_picture, op_pix, op_qpix);
1063 /* add dct residue */
1064 add_dct(s, block[0], 0, dest_y, dct_linesize);
1065 add_dct(s, block[1], 1, dest_y + 8, dct_linesize);
1066 add_dct(s, block[2], 2, dest_y + dct_offset, dct_linesize);
1067 add_dct(s, block[3], 3, dest_y + dct_offset + 8, dct_linesize);
1069 add_dct(s, block[4], 4, dest_cb, s->linesize >> 1);
1070 add_dct(s, block[5], 5, dest_cr, s->linesize >> 1);
1072 /* dct only in intra block */
1073 put_dct(s, block[0], 0, dest_y, dct_linesize);
1074 put_dct(s, block[1], 1, dest_y + 8, dct_linesize);
1075 put_dct(s, block[2], 2, dest_y + dct_offset, dct_linesize);
1076 put_dct(s, block[3], 3, dest_y + dct_offset + 8, dct_linesize);
1078 put_dct(s, block[4], 4, dest_cb, s->linesize >> 1);
1079 put_dct(s, block[5], 5, dest_cr, s->linesize >> 1);
1083 emms_c(); //FIXME remove
1086 static void encode_mb(MpegEncContext *s)
1089 const int mb_x= s->mb_x;
1090 const int mb_y= s->mb_y;
1092 const int motion_x= s->mv[0][0][0];
1093 const int motion_y= s->mv[0][0][1];
1096 /* get the pixels */
1098 ptr = s->new_picture[0] + (mb_y * 16 * wrap) + mb_x * 16;
1099 get_pixels(s->block[0], ptr, wrap);
1100 get_pixels(s->block[1], ptr + 8, wrap);
1101 get_pixels(s->block[2], ptr + 8 * wrap, wrap);
1102 get_pixels(s->block[3], ptr + 8 * wrap + 8, wrap);
1103 wrap = s->linesize >> 1;
1104 ptr = s->new_picture[1] + (mb_y * 8 * wrap) + mb_x * 8;
1105 get_pixels(s->block[4], ptr, wrap);
1107 wrap = s->linesize >> 1;
1108 ptr = s->new_picture[2] + (mb_y * 8 * wrap) + mb_x * 8;
1109 get_pixels(s->block[5], ptr, wrap);
1111 /* subtract previous frame if non intra */
1113 int dxy, offset, mx, my;
1115 dxy = ((motion_y & 1) << 1) | (motion_x & 1);
1116 ptr = s->last_picture[0] +
1117 ((mb_y * 16 + (motion_y >> 1)) * s->linesize) +
1118 (mb_x * 16 + (motion_x >> 1));
1120 sub_pixels_2(s->block[0], ptr, s->linesize, dxy);
1121 sub_pixels_2(s->block[1], ptr + 8, s->linesize, dxy);
1122 sub_pixels_2(s->block[2], ptr + s->linesize * 8, s->linesize, dxy);
1123 sub_pixels_2(s->block[3], ptr + 8 + s->linesize * 8, s->linesize ,dxy);
1125 if (s->out_format == FMT_H263) {
1126 /* special rounding for h263 */
1128 if ((motion_x & 3) != 0)
1130 if ((motion_y & 3) != 0)
1137 dxy = ((my & 1) << 1) | (mx & 1);
1141 offset = ((mb_y * 8 + my) * (s->linesize >> 1)) + (mb_x * 8 + mx);
1142 ptr = s->last_picture[1] + offset;
1143 sub_pixels_2(s->block[4], ptr, s->linesize >> 1, dxy);
1144 ptr = s->last_picture[2] + offset;
1145 sub_pixels_2(s->block[5], ptr, s->linesize >> 1, dxy);
1152 adap_parm = ((s->avg_mb_var << 1) + s->mb_var[s->mb_width*mb_y+mb_x] + 1.0) /
1153 ((s->mb_var[s->mb_width*mb_y+mb_x] << 1) + s->avg_mb_var + 1.0);
1155 printf("\ntype=%c qscale=%2d adap=%0.2f dquant=%4.2f var=%4d avgvar=%4d",
1156 (s->mb_type[s->mb_width*mb_y+mb_x] > 0) ? 'I' : 'P',
1157 s->qscale, adap_parm, s->qscale*adap_parm,
1158 s->mb_var[s->mb_width*mb_y+mb_x], s->avg_mb_var);
1161 /* DCT & quantize */
1162 if (s->h263_msmpeg4) {
1163 msmpeg4_dc_scale(s);
1164 } else if (s->h263_pred) {
1167 /* default quantization values */
1172 s->block_last_index[i] = dct_quantize(s, s->block[i], i, s->qscale);
1175 /* huffman encode */
1176 switch(s->out_format) {
1178 mpeg1_encode_mb(s, s->block, motion_x, motion_y);
1181 if (s->h263_msmpeg4)
1182 msmpeg4_encode_mb(s, s->block, motion_x, motion_y);
1183 else if(s->h263_pred)
1184 mpeg4_encode_mb(s, s->block, motion_x, motion_y);
1186 h263_encode_mb(s, s->block, motion_x, motion_y);
1189 mjpeg_encode_mb(s, s->block);
1194 static void copy_bits(PutBitContext *pb, UINT8 *src, int length)
1196 int bytes= length>>3;
1200 for(i=0; i<bytes; i++) put_bits(pb, 8, src[i]);
1201 put_bits(pb, bits, src[i]>>(8-bits));
1204 static void encode_picture(MpegEncContext *s, int picture_number)
1206 int mb_x, mb_y, last_gob, pdif = 0;
1209 MpegEncContext best_s;
1210 UINT8 bit_buf[4][3000]; //FIXME check that this is ALLWAYS large enogh for a MB
1212 s->picture_number = picture_number;
1217 s->block_wrap[3]= s->mb_width*2 + 2;
1219 s->block_wrap[5]= s->mb_width + 2;
1221 s->last_mc_mb_var = s->mc_mb_var;
1222 /* Reset the average MB variance */
1225 /* Estimate motion for every MB */
1226 if(s->pict_type == P_TYPE){
1227 for(mb_y=0; mb_y < s->mb_height; mb_y++) {
1228 s->block_index[0]= s->block_wrap[0]*(mb_y*2 + 1) - 1;
1229 s->block_index[1]= s->block_wrap[0]*(mb_y*2 + 1);
1230 s->block_index[2]= s->block_wrap[0]*(mb_y*2 + 2) - 1;
1231 s->block_index[3]= s->block_wrap[0]*(mb_y*2 + 2);
1232 for(mb_x=0; mb_x < s->mb_width; mb_x++) {
1235 s->block_index[0]+=2;
1236 s->block_index[1]+=2;
1237 s->block_index[2]+=2;
1238 s->block_index[3]+=2;
1240 /* compute motion vector & mb_type and store in context */
1241 estimate_motion(s, mb_x, mb_y);
1242 // s->mb_type[mb_y*s->mb_width + mb_x]=MB_TYPE_INTER;
1248 //FIXME do we need to zero them?
1249 memset(s->motion_val[0], 0, sizeof(INT16)*(s->mb_width*2 + 2)*(s->mb_height*2 + 2)*2);
1250 memset(s->mv_table[0] , 0, sizeof(INT16)*s->mb_width*s->mb_height);
1251 memset(s->mv_table[1] , 0, sizeof(INT16)*s->mb_width*s->mb_height);
1252 memset(s->mb_type , MB_TYPE_INTRA, sizeof(UINT8)*s->mb_width*s->mb_height);
1255 if(s->avg_mb_var < s->mc_mb_var && s->pict_type != B_TYPE){ //FIXME subtract MV bits
1256 s->pict_type= I_TYPE;
1257 s->picture_in_gop_number=0;
1258 memset(s->mb_type , MB_TYPE_INTRA, sizeof(UINT8)*s->mb_width*s->mb_height);
1259 //printf("Scene change detected, encoding as I Frame\n");
1262 /* find best f_code for ME which do unlimited searches */
1263 if(s->pict_type==P_TYPE && s->full_search>3){
1267 UINT8 * fcode_tab= s->fcode_tab;
1269 for(i=0; i<8; i++) mv_num[i]=0;
1271 for(i=0; i<s->mb_num; i++){
1272 if(s->mb_type[i] & (MB_TYPE_INTER|MB_TYPE_INTER4V)){
1273 mv_num[ fcode_tab[s->mv_table[0][i] + MAX_MV] ]++;
1274 mv_num[ fcode_tab[s->mv_table[1][i] + MAX_MV] ]++;
1275 //printf("%d %d %d\n", s->mv_table[0][i], fcode_tab[s->mv_table[0][i] + MAX_MV], i);
1280 for(i=MAX_FCODE; i>1; i--){
1282 if(loose > 10) break; //FIXME this is pretty ineffective
1285 /* for(i=0; i<=MAX_FCODE; i++){
1286 printf("%d ", mv_num[i]);
1293 //printf("f_code %d ///\n", s->f_code);
1294 /* convert MBs with too long MVs to I-Blocks */
1295 if(s->pict_type==P_TYPE){
1297 const int f_code= s->f_code;
1298 UINT8 * fcode_tab= s->fcode_tab;
1300 for(i=0; i<s->mb_num; i++){
1301 if(s->mb_type[i]&MB_TYPE_INTER){
1302 if( fcode_tab[s->mv_table[0][i] + MAX_MV] > f_code
1303 || fcode_tab[s->mv_table[0][i] + MAX_MV] == 0
1304 || fcode_tab[s->mv_table[1][i] + MAX_MV] > f_code
1305 || fcode_tab[s->mv_table[1][i] + MAX_MV] == 0 ){
1306 s->mb_type[i] &= ~MB_TYPE_INTER;
1307 s->mb_type[i] |= MB_TYPE_INTRA;
1308 s->mv_table[0][i] = 0;
1309 s->mv_table[1][i] = 0;
1312 if(s->mb_type[i]&MB_TYPE_INTER4V){
1318 // printf("%d %d\n", s->avg_mb_var, s->mc_mb_var);
1320 if (!s->fixed_qscale)
1321 s->qscale = rate_estimate_qscale(s);
1323 /* precompute matrix */
1324 if (s->out_format == FMT_MJPEG) {
1325 /* for mjpeg, we do include qscale in the matrix */
1326 s->intra_matrix[0] = default_intra_matrix[0];
1328 s->intra_matrix[i] = (default_intra_matrix[i] * s->qscale) >> 3;
1329 convert_matrix(s->q_intra_matrix, s->q_intra_matrix16, s->intra_matrix, 8);
1331 convert_matrix(s->q_intra_matrix, s->q_intra_matrix16, s->intra_matrix, s->qscale);
1332 convert_matrix(s->q_non_intra_matrix, s->q_non_intra_matrix16, s->non_intra_matrix, s->qscale);
1335 s->last_bits= get_bit_count(&s->pb);
1336 switch(s->out_format) {
1338 mjpeg_picture_header(s);
1341 if (s->h263_msmpeg4)
1342 msmpeg4_encode_picture_header(s, picture_number);
1343 else if (s->h263_pred)
1344 mpeg4_encode_picture_header(s, picture_number);
1345 else if (s->h263_rv10)
1346 rv10_encode_picture_header(s, picture_number);
1348 h263_encode_picture_header(s, picture_number);
1351 mpeg1_encode_picture_header(s, picture_number);
1354 bits= get_bit_count(&s->pb);
1355 s->header_bits= bits - s->last_bits;
1365 /* init last dc values */
1366 /* note: quant matrix value (8) is implied here */
1367 s->last_dc[0] = 128;
1368 s->last_dc[1] = 128;
1369 s->last_dc[2] = 128;
1371 s->last_mv[0][0][0] = 0;
1372 s->last_mv[0][0][1] = 0;
1374 /* Get the GOB height based on picture height */
1375 if (s->out_format == FMT_H263 && !s->h263_pred && !s->h263_msmpeg4) {
1376 if (s->height <= 400)
1378 else if (s->height <= 800)
1384 s->avg_mb_var = s->avg_mb_var / s->mb_num;
1386 for(mb_y=0; mb_y < s->mb_height; mb_y++) {
1387 /* Put GOB header based on RTP MTU */
1388 /* TODO: Put all this stuff in a separate generic function */
1391 s->ptr_lastgob = s->pb.buf;
1392 s->ptr_last_mb_line = s->pb.buf;
1393 } else if (s->out_format == FMT_H263 && !s->h263_pred && !s->h263_msmpeg4 && !(mb_y % s->gob_index)) {
1394 last_gob = h263_encode_gob_header(s, mb_y);
1396 s->first_gob_line = 1;
1401 s->block_index[0]= s->block_wrap[0]*(mb_y*2 + 1) - 1;
1402 s->block_index[1]= s->block_wrap[0]*(mb_y*2 + 1);
1403 s->block_index[2]= s->block_wrap[0]*(mb_y*2 + 2) - 1;
1404 s->block_index[3]= s->block_wrap[0]*(mb_y*2 + 2);
1405 s->block_index[4]= s->block_wrap[4]*(mb_y + 1) + s->block_wrap[0]*(s->mb_height*2 + 2);
1406 s->block_index[5]= s->block_wrap[4]*(mb_y + 1 + s->mb_height + 2) + s->block_wrap[0]*(s->mb_height*2 + 2);
1407 for(mb_x=0; mb_x < s->mb_width; mb_x++) {
1408 const int mb_type= s->mb_type[mb_y * s->mb_width + mb_x];
1416 s->block_index[0]+=2;
1417 s->block_index[1]+=2;
1418 s->block_index[2]+=2;
1419 s->block_index[3]+=2;
1420 s->block_index[4]++;
1421 s->block_index[5]++;
1423 s->mv_type = MV_TYPE_16X16;
1424 s->mv_dir = MV_DIR_FORWARD;
1425 if(mb_type & (mb_type-1)){ // more than 1 MB type possible
1427 if(mb_type&MB_TYPE_INTER){
1429 s->mv[0][0][0] = s->mv_table[0][mb_y * s->mb_width + mb_x];
1430 s->mv[0][0][1] = s->mv_table[1][mb_y * s->mb_width + mb_x];
1431 init_put_bits(&s->pb, bit_buf[1], 3000, NULL, NULL);
1432 s->block= s->inter_block;
1435 d= get_bit_count(&s->pb);
1437 flush_put_bits(&s->pb);
1439 best_s.mv[0][0][0]= s->mv[0][0][0];
1440 best_s.mv[0][0][1]= s->mv[0][0][1];
1443 best_s.block= s->block;
1446 best_s.block_last_index[i]= s->block_last_index[i];
1449 if(mb_type&MB_TYPE_INTRA){
1453 init_put_bits(&s->pb, bit_buf[0], 3000, NULL, NULL);
1454 s->block= s->intra_block;
1457 d= get_bit_count(&s->pb);
1459 flush_put_bits(&s->pb);
1461 best_s.mv[0][0][0]= 0;
1462 best_s.mv[0][0][1]= 0;
1465 best_s.block= s->block;
1467 best_s.block_last_index[i]= s->block_last_index[i];
1470 /* force cleaning of ac/dc if needed ... */
1471 s->mbintra_table[mb_x + mb_y*s->mb_width]=1;
1473 s->mv[0][0][0]= best_s.mv[0][0][0];
1474 s->mv[0][0][1]= best_s.mv[0][0][1];
1475 s->mb_intra= best_s.mb_intra;
1477 s->block_last_index[i]= best_s.block_last_index[i];
1478 copy_bits(&pb, bit_buf[best], dmin);
1479 s->block= best_s.block;
1482 // only one MB-Type possible
1483 if(mb_type&MB_TYPE_INTRA){
1489 s->mv[0][0][0] = s->mv_table[0][mb_y * s->mb_width + mb_x];
1490 s->mv[0][0][1] = s->mv_table[1][mb_y * s->mb_width + mb_x];
1495 MPV_decode_mb(s, s->block);
1499 /* Obtain average GOB size for RTP */
1502 s->mb_line_avgsize = pbBufPtr(&s->pb) - s->ptr_last_mb_line;
1503 else if (!(mb_y % s->gob_index)) {
1504 s->mb_line_avgsize = (s->mb_line_avgsize + pbBufPtr(&s->pb) - s->ptr_last_mb_line) >> 1;
1505 s->ptr_last_mb_line = pbBufPtr(&s->pb);
1507 //fprintf(stderr, "\nMB line: %d\tSize: %u\tAvg. Size: %u", s->mb_y,
1508 // (s->pb.buf_ptr - s->ptr_last_mb_line), s->mb_line_avgsize);
1509 s->first_gob_line = 0;
1514 if (s->h263_msmpeg4 && s->pict_type == I_TYPE)
1515 msmpeg4_encode_ext_header(s);
1517 //if (s->gob_number)
1518 // fprintf(stderr,"\nNumber of GOB: %d", s->gob_number);
1520 /* Send the last GOB if RTP */
1522 flush_put_bits(&s->pb);
1523 pdif = pbBufPtr(&s->pb) - s->ptr_lastgob;
1524 /* Call the RTP callback to send the last GOB */
1525 if (s->rtp_callback)
1526 s->rtp_callback(s->ptr_lastgob, pdif, s->gob_number);
1527 s->ptr_lastgob = pbBufPtr(&s->pb);
1528 //fprintf(stderr,"\nGOB: %2d size: %d (last)", s->gob_number, pdif);
1532 static int dct_quantize_c(MpegEncContext *s,
1533 DCTELEM *block, int n,
1536 int i, j, level, last_non_zero, q;
1538 int minLevel, maxLevel;
1540 if(s->avctx!=NULL && s->avctx->codec->id==CODEC_ID_MPEG4){
1544 }else if(s->out_format==FMT_MPEG1){
1548 }else if(s->out_format==FMT_MJPEG){
1553 /* h263 / msmpeg4 */
1560 /* we need this permutation so that we correct the IDCT
1561 permutation. will be moved into DCT code */
1562 block_permute(block);
1571 /* note: block[0] is assumed to be positive */
1572 block[0] = (block[0] + (q >> 1)) / q;
1575 if (s->out_format == FMT_H263) {
1576 qmat = s->q_non_intra_matrix;
1578 qmat = s->q_intra_matrix;
1583 qmat = s->q_non_intra_matrix;
1587 j = zigzag_direct[i];
1589 level = level * qmat[j];
1592 static int count = 0;
1593 int level1, level2, qmat1;
1595 if (qmat == s->q_non_intra_matrix) {
1596 qmat1 = default_non_intra_matrix[j] * s->qscale;
1598 qmat1 = default_intra_matrix[j] * s->qscale;
1600 if (av_fdct != jpeg_fdct_ifast)
1601 val = ((double)block[j] * 8.0) / (double)qmat1;
1603 val = ((double)block[j] * 8.0 * 2048.0) /
1604 ((double)qmat1 * aanscales[j]);
1606 level2 = level / (1 << (QMAT_SHIFT - 3));
1607 if (level1 != level2) {
1608 fprintf(stderr, "%d: quant error qlevel=%d wanted=%d level=%d qmat1=%d qmat=%d wantedf=%0.6f\n",
1609 count, level2, level1, block[j], qmat1, qmat[j],
1616 /* XXX: slight error for the low range. Test should be equivalent to
1617 (level <= -(1 << (QMAT_SHIFT - 3)) || level >= (1 <<
1620 if (((level << (31 - (QMAT_SHIFT - 3))) >> (31 - (QMAT_SHIFT - 3))) !=
1622 level = level / (1 << (QMAT_SHIFT - 3));
1623 /* XXX: currently, this code is not optimal. the range should be:
1629 if (level > maxLevel)
1631 else if (level < minLevel)
1640 return last_non_zero;
1643 static void dct_unquantize_mpeg1_c(MpegEncContext *s,
1644 DCTELEM *block, int n, int qscale)
1646 int i, level, nCoeffs;
1647 const UINT16 *quant_matrix;
1649 if(s->alternate_scan) nCoeffs= 64;
1650 else nCoeffs= s->block_last_index[n]+1;
1654 block[0] = block[0] * s->y_dc_scale;
1656 block[0] = block[0] * s->c_dc_scale;
1657 /* XXX: only mpeg1 */
1658 quant_matrix = s->intra_matrix;
1659 for(i=1;i<nCoeffs;i++) {
1660 int j= zigzag_direct[i];
1665 level = (int)(level * qscale * quant_matrix[j]) >> 3;
1666 level = (level - 1) | 1;
1669 level = (int)(level * qscale * quant_matrix[j]) >> 3;
1670 level = (level - 1) | 1;
1673 if (level < -2048 || level > 2047)
1674 fprintf(stderr, "unquant error %d %d\n", i, level);
1681 quant_matrix = s->non_intra_matrix;
1682 for(;i<nCoeffs;i++) {
1683 int j= zigzag_direct[i];
1688 level = (((level << 1) + 1) * qscale *
1689 ((int) (quant_matrix[j]))) >> 4;
1690 level = (level - 1) | 1;
1693 level = (((level << 1) + 1) * qscale *
1694 ((int) (quant_matrix[j]))) >> 4;
1695 level = (level - 1) | 1;
1698 if (level < -2048 || level > 2047)
1699 fprintf(stderr, "unquant error %d %d\n", i, level);
1707 static void dct_unquantize_h263_c(MpegEncContext *s,
1708 DCTELEM *block, int n, int qscale)
1710 int i, level, qmul, qadd;
1716 block[0] = block[0] * s->y_dc_scale;
1718 block[0] = block[0] * s->c_dc_scale;
1721 nCoeffs= 64; //does not allways use zigzag table
1724 nCoeffs= zigzag_end[ s->block_last_index[n] ];
1727 qmul = s->qscale << 1;
1728 if (s->h263_aic && s->mb_intra)
1731 qadd = (s->qscale - 1) | 1;
1733 for(;i<nCoeffs;i++) {
1737 level = level * qmul - qadd;
1739 level = level * qmul + qadd;
1742 if (level < -2048 || level > 2047)
1743 fprintf(stderr, "unquant error %d %d\n", i, level);
1752 /* an I frame is I_FRAME_SIZE_RATIO bigger than a P frame */
1753 #define I_FRAME_SIZE_RATIO 3.0
1756 static void rate_control_init(MpegEncContext *s)
1761 //initial values, they dont really matter as they will be totally different within a few frames
1762 s->i_pred.coeff= s->p_pred.coeff= 7.0;
1763 s->i_pred.count= s->p_pred.count= 1.0;
1765 s->i_pred.decay= s->p_pred.decay= 0.4;
1767 // use more bits at the beginning, otherwise high motion at the begin will look like shit
1771 s->short_term_qsum=0.001;
1772 s->short_term_qcount=0.001;
1776 if (s->intra_only) {
1777 s->I_frame_bits = ((INT64)s->bit_rate * FRAME_RATE_BASE) / s->frame_rate;
1778 s->P_frame_bits = s->I_frame_bits;
1780 s->P_frame_bits = (int) ((float)(s->gop_size * s->bit_rate) /
1781 (float)((float)s->frame_rate / FRAME_RATE_BASE * (I_FRAME_SIZE_RATIO + s->gop_size - 1)));
1782 s->I_frame_bits = (int)(s->P_frame_bits * I_FRAME_SIZE_RATIO);
1786 printf("I_frame_size=%d P_frame_size=%d\n",
1787 s->I_frame_bits, s->P_frame_bits);
1792 static double predict(Predictor *p, double q, double var)
1794 return p->coeff*var / (q*p->count);
1797 static void update_predictor(Predictor *p, double q, double var, double size)
1799 double new_coeff= size*q / (var + 1);
1800 if(var<1000) return;
1802 int pred= predict(p, q, var);
1803 int error= abs(pred-size);
1804 static double sum=0;
1806 if(count>5) sum+=error;
1808 if(256*256*256*64%count==0){
1809 printf("%d %f %f\n", count, sum/count, p->coeff);
1812 p->count*= p->decay;
1813 p->coeff*= p->decay;
1815 p->coeff+= new_coeff;
1818 static int rate_estimate_qscale(MpegEncContext *s)
1826 float br_compensation;
1828 double short_term_q;
1830 int last_qscale= s->qscale;
1835 fps= (double)s->frame_rate / FRAME_RATE_BASE;
1836 wanted_bits= s->bit_rate*(double)s->picture_number/fps;
1839 if(s->picture_number>2){
1840 /* update predictors */
1841 if(s->last_pict_type == I_TYPE){
1844 //printf("%d %d %d %f\n", s->qscale, s->last_mc_mb_var, s->frame_bits, s->p_pred.coeff);
1845 update_predictor(&s->p_pred, s->qscale, s->last_mc_mb_var, s->frame_bits);
1849 if(s->pict_type == I_TYPE){
1851 rate_q= s->qsum/s->qcount;
1854 int diff, best_diff=1000000000;
1855 for(i=1; i<=31; i++){
1856 diff= predict(&s->p_pred, i, s->mc_mb_var) - (double)s->bit_rate/fps;
1857 if(diff<0) diff= -diff;
1865 s->short_term_qsum*=s->qblur;
1866 s->short_term_qcount*=s->qblur;
1868 s->short_term_qsum+= rate_q;
1869 s->short_term_qcount++;
1870 short_term_q= s->short_term_qsum/s->short_term_qcount;
1872 long_term_q= s->qsum/s->qcount*s->total_bits/wanted_bits;
1874 // q= (long_term_q - short_term_q)*s->qcompress + short_term_q;
1875 q= 1/((1/long_term_q - 1/short_term_q)*s->qcompress + 1/short_term_q);
1877 diff= s->total_bits - wanted_bits;
1878 br_compensation= (s->bit_rate_tolerance - diff)/s->bit_rate_tolerance;
1879 if(br_compensation<=0.0) br_compensation=0.001;
1882 qscale= (int)(q + 0.5);
1883 if (qscale<qmin) qscale=qmin;
1884 else if(qscale>qmax) qscale=qmax;
1886 if (qscale<last_qscale-s->max_qdiff) qscale=last_qscale-s->max_qdiff;
1887 else if(qscale>last_qscale+s->max_qdiff) qscale=last_qscale+s->max_qdiff;
1892 s->last_pict_type= s->pict_type;
1893 //printf("q:%d diff:%d comp:%f rate_q:%d st_q:%d fvar:%d last_size:%d\n", qscale, (int)diff, br_compensation,
1894 // rate_q, (int)short_term_q, s->mc_mb_var, s->frame_bits);
1895 //printf("%d %d\n", s->bit_rate, (int)fps);
1898 INT64 diff, total_bits = s->total_bits;
1901 if (s->pict_type == I_TYPE) {
1902 s->wanted_bits += s->I_frame_bits;
1904 s->wanted_bits += s->P_frame_bits;
1906 diff = s->wanted_bits - total_bits;
1907 q = 31.0 - (float)diff / (QSCALE_K * s->mb_height * s->mb_width);
1908 /* adjust for I frame */
1909 if (s->pict_type == I_TYPE && !s->intra_only) {
1910 q /= I_FRAME_SIZE_RATIO;
1913 /* using a too small Q scale leeds to problems in mpeg1 and h263
1914 because AC coefficients are clamped to 255 or 127 */
1920 qscale = (int)(q + 0.5);
1922 printf("\n%d: total=%0.0f wanted=%0.0f br=%0.1f diff=%d qest=%2.1f\n",
1925 (double)s->wanted_bits,
1926 (float)s->frame_rate / FRAME_RATE_BASE *
1927 total_bits / s->picture_number,
1934 AVCodec mpeg1video_encoder = {
1937 CODEC_ID_MPEG1VIDEO,
1938 sizeof(MpegEncContext),
1944 AVCodec h263_encoder = {
1948 sizeof(MpegEncContext),
1954 AVCodec h263p_encoder = {
1958 sizeof(MpegEncContext),
1964 AVCodec rv10_encoder = {
1968 sizeof(MpegEncContext),
1974 AVCodec mjpeg_encoder = {
1978 sizeof(MpegEncContext),
1984 AVCodec mpeg4_encoder = {
1988 sizeof(MpegEncContext),
1994 AVCodec msmpeg4_encoder = {
1998 sizeof(MpegEncContext),