ring buffer fix by (Luca Abeni <lucabe72 at email dot it>)
[ffmpeg.git] / libavformat / grab.c
1 /*
2  * Linux video grab interface
3  * Copyright (c) 2000,2001 Fabrice Bellard.
4  *
5  * This library is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU Lesser General Public
7  * License as published by the Free Software Foundation; either
8  * version 2 of the License, or (at your option) any later version.
9  *
10  * This library 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 GNU
13  * Lesser General Public License for more details.
14  *
15  * You should have received a copy of the GNU Lesser General Public
16  * License along with this library; if not, write to the Free Software
17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
18  */
19 #include "avformat.h"
20 #include <unistd.h>
21 #include <fcntl.h>
22 #include <sys/ioctl.h>
23 #include <sys/mman.h>
24 #include <sys/time.h>
25 #define _LINUX_TIME_H 1
26 #include <linux/videodev.h>
27 #include <time.h>
28
29 typedef struct {
30     int fd;
31     int frame_format; /* see VIDEO_PALETTE_xxx */
32     int use_mmap;
33     int width, height;
34     int frame_rate;
35     int frame_rate_base;
36     int64_t time_frame;
37     int frame_size;
38     struct video_capability video_cap;
39     struct video_audio audio_saved;
40     uint8_t *video_buf;
41     struct video_mbuf gb_buffers;
42     struct video_mmap gb_buf;
43     int gb_frame;
44
45     /* ATI All In Wonder specific stuff */
46     /* XXX: remove and merge in libavcodec/imgconvert.c */
47     int aiw_enabled;
48     int deint;
49     int halfw;
50     uint8_t *src_mem;
51     uint8_t *lum_m4_mem;
52 } VideoData;
53
54 static int aiw_init(VideoData *s);
55 static int aiw_read_picture(VideoData *s, uint8_t *data);
56 static int aiw_close(VideoData *s);
57
58 static int grab_read_header(AVFormatContext *s1, AVFormatParameters *ap)
59 {
60     VideoData *s = s1->priv_data;
61     AVStream *st;
62     int width, height;
63     int video_fd, frame_size;
64     int ret, frame_rate, frame_rate_base;
65     int desired_palette;
66     struct video_tuner tuner;
67     struct video_audio audio;
68     const char *video_device;
69     int j;
70
71     if (!ap || ap->width <= 0 || ap->height <= 0 || ap->frame_rate <= 0)
72         return -1;
73     
74     width = ap->width;
75     height = ap->height;
76     frame_rate      = ap->frame_rate;
77     frame_rate_base = ap->frame_rate_base;
78
79     st = av_new_stream(s1, 0);
80     if (!st)
81         return -ENOMEM;
82
83     s->width = width;
84     s->height = height;
85     s->frame_rate      = frame_rate;
86     s->frame_rate_base = frame_rate_base;
87
88     video_device = ap->device;
89     if (!video_device)
90         video_device = "/dev/video";
91     video_fd = open(video_device, O_RDWR);
92     if (video_fd < 0) {
93         perror(video_device);
94         goto fail;
95     }
96     
97     if (ioctl(video_fd,VIDIOCGCAP, &s->video_cap) < 0) {
98         perror("VIDIOCGCAP");
99         goto fail;
100     }
101
102     if (!(s->video_cap.type & VID_TYPE_CAPTURE)) {
103         av_log(s1, AV_LOG_ERROR, "Fatal: grab device does not handle capture\n");
104         goto fail;
105     }
106
107     desired_palette = -1;
108     if (st->codec.pix_fmt == PIX_FMT_YUV420P) {
109         desired_palette = VIDEO_PALETTE_YUV420P;
110     } else if (st->codec.pix_fmt == PIX_FMT_YUV422) {
111         desired_palette = VIDEO_PALETTE_YUV422;
112     } else if (st->codec.pix_fmt == PIX_FMT_BGR24) {
113         desired_palette = VIDEO_PALETTE_RGB24;
114     }    
115
116     /* set tv standard */
117     if (ap->standard && !ioctl(video_fd, VIDIOCGTUNER, &tuner)) {
118         if (!strcasecmp(ap->standard, "pal"))
119             tuner.mode = VIDEO_MODE_PAL;
120         else if (!strcasecmp(ap->standard, "secam"))
121             tuner.mode = VIDEO_MODE_SECAM;
122         else
123             tuner.mode = VIDEO_MODE_NTSC;
124         ioctl(video_fd, VIDIOCSTUNER, &tuner);
125     }
126     
127     /* unmute audio */
128     audio.audio = 0;
129     ioctl(video_fd, VIDIOCGAUDIO, &audio);
130     memcpy(&s->audio_saved, &audio, sizeof(audio));
131     audio.flags &= ~VIDEO_AUDIO_MUTE;
132     ioctl(video_fd, VIDIOCSAUDIO, &audio);
133
134     ret = ioctl(video_fd,VIDIOCGMBUF,&s->gb_buffers);
135     if (ret < 0) {
136         /* try to use read based access */
137         struct video_window win;
138         struct video_picture pict;
139         int val;
140
141         win.x = 0;
142         win.y = 0;
143         win.width = width;
144         win.height = height;
145         win.chromakey = -1;
146         win.flags = 0;
147
148         ioctl(video_fd, VIDIOCSWIN, &win);
149
150         ioctl(video_fd, VIDIOCGPICT, &pict);
151 #if 0
152         printf("v4l: colour=%d hue=%d brightness=%d constrast=%d whiteness=%d\n",
153                pict.colour,
154                pict.hue,
155                pict.brightness,
156                pict.contrast,
157                pict.whiteness);
158 #endif        
159         /* try to choose a suitable video format */
160         pict.palette = desired_palette;
161         if (desired_palette == -1 || (ret = ioctl(video_fd, VIDIOCSPICT, &pict)) < 0) {
162             pict.palette=VIDEO_PALETTE_YUV420P;
163             ret = ioctl(video_fd, VIDIOCSPICT, &pict);
164             if (ret < 0) {
165                 pict.palette=VIDEO_PALETTE_YUV422;
166                 ret = ioctl(video_fd, VIDIOCSPICT, &pict);
167                 if (ret < 0) {
168                     pict.palette=VIDEO_PALETTE_RGB24;
169                     ret = ioctl(video_fd, VIDIOCSPICT, &pict);
170                     if (ret < 0) 
171                         goto fail1;
172                 }
173             }
174         }
175
176         s->frame_format = pict.palette;
177
178         val = 1;
179         ioctl(video_fd, VIDIOCCAPTURE, &val);
180
181         s->time_frame = av_gettime();
182         s->use_mmap = 0;
183         
184         /* ATI All In Wonder automatic activation */
185         if (!strcmp(s->video_cap.name, "Km")) {
186             if (aiw_init(s) < 0)
187                 goto fail;
188             s->aiw_enabled = 1;
189             /* force 420P format because convertion from YUV422 to YUV420P
190                is done in this driver (ugly) */
191             s->frame_format = VIDEO_PALETTE_YUV420P;
192         }
193     } else {
194         s->video_buf = mmap(0,s->gb_buffers.size,PROT_READ|PROT_WRITE,MAP_SHARED,video_fd,0);
195         if ((unsigned char*)-1 == s->video_buf) {
196             perror("mmap");
197             goto fail;
198         }
199         s->gb_frame = 0;
200         s->time_frame = av_gettime();
201         
202         /* start to grab the first frame */
203         s->gb_buf.frame = s->gb_frame % s->gb_buffers.frames;
204         s->gb_buf.height = height;
205         s->gb_buf.width = width;
206         s->gb_buf.format = desired_palette;
207
208         if (desired_palette == -1 || (ret = ioctl(video_fd, VIDIOCMCAPTURE, &s->gb_buf)) < 0) {
209             s->gb_buf.format = VIDEO_PALETTE_YUV420P;
210             
211             ret = ioctl(video_fd, VIDIOCMCAPTURE, &s->gb_buf);
212             if (ret < 0 && errno != EAGAIN) {
213                 /* try YUV422 */
214                 s->gb_buf.format = VIDEO_PALETTE_YUV422;
215                 
216                 ret = ioctl(video_fd, VIDIOCMCAPTURE, &s->gb_buf);
217                 if (ret < 0 && errno != EAGAIN) {
218                     /* try RGB24 */
219                     s->gb_buf.format = VIDEO_PALETTE_RGB24;
220                     ret = ioctl(video_fd, VIDIOCMCAPTURE, &s->gb_buf);
221                 }
222             }
223         }
224         if (ret < 0) {
225             if (errno != EAGAIN) {
226             fail1:
227                 av_log(s1, AV_LOG_ERROR, "Fatal: grab device does not support suitable format\n");
228             } else {
229                 av_log(s1, AV_LOG_ERROR,"Fatal: grab device does not receive any video signal\n");
230             }
231             goto fail;
232         }
233         for (j = 1; j < s->gb_buffers.frames; j++) {
234           s->gb_buf.frame = j;
235           ioctl(video_fd, VIDIOCMCAPTURE, &s->gb_buf);
236         }
237         s->frame_format = s->gb_buf.format;
238         s->use_mmap = 1;
239     }
240
241     switch(s->frame_format) {
242     case VIDEO_PALETTE_YUV420P:
243         frame_size = (width * height * 3) / 2;
244         st->codec.pix_fmt = PIX_FMT_YUV420P;
245         break;
246     case VIDEO_PALETTE_YUV422:
247         frame_size = width * height * 2;
248         st->codec.pix_fmt = PIX_FMT_YUV422;
249         break;
250     case VIDEO_PALETTE_RGB24:
251         frame_size = width * height * 3;
252         st->codec.pix_fmt = PIX_FMT_BGR24; /* NOTE: v4l uses BGR24, not RGB24 ! */
253         break;
254     default:
255         goto fail;
256     }
257     s->fd = video_fd;
258     s->frame_size = frame_size;
259     
260     st->codec.codec_type = CODEC_TYPE_VIDEO;
261     st->codec.codec_id = CODEC_ID_RAWVIDEO;
262     st->codec.width = width;
263     st->codec.height = height;
264     st->codec.frame_rate      = frame_rate;
265     st->codec.frame_rate_base = frame_rate_base;
266     
267     av_set_pts_info(s1, 48, 1, 1000000); /* 48 bits pts in us */
268
269     return 0;
270  fail:
271     if (video_fd >= 0)
272         close(video_fd);
273     av_free(st);
274     return -EIO;
275 }
276
277 static int v4l_mm_read_picture(VideoData *s, uint8_t *buf)
278 {
279     uint8_t *ptr;
280
281     while (ioctl(s->fd, VIDIOCSYNC, &s->gb_frame) < 0 &&
282            (errno == EAGAIN || errno == EINTR));
283
284     ptr = s->video_buf + s->gb_buffers.offsets[s->gb_frame];
285     memcpy(buf, ptr, s->frame_size);
286
287     /* Setup to capture the next frame */
288     s->gb_buf.frame = s->gb_frame;
289     if (ioctl(s->fd, VIDIOCMCAPTURE, &s->gb_buf) < 0) {
290         if (errno == EAGAIN)
291             av_log(NULL, AV_LOG_ERROR, "Cannot Sync\n");
292         else
293             perror("VIDIOCMCAPTURE");
294         return -EIO;
295     }
296
297     /* This is now the grabbing frame */
298     s->gb_frame = (s->gb_frame + 1) % s->gb_buffers.frames;
299
300     return s->frame_size;
301 }
302
303 static int grab_read_packet(AVFormatContext *s1, AVPacket *pkt)
304 {
305     VideoData *s = s1->priv_data;
306     int64_t curtime, delay;
307     struct timespec ts;
308     int64_t per_frame = (int64_t_C(1000000) * s->frame_rate_base) / s->frame_rate;
309
310     /* Calculate the time of the next frame */
311     s->time_frame += per_frame;
312
313     /* wait based on the frame rate */
314     for(;;) {
315         curtime = av_gettime();
316         delay = s->time_frame - curtime;
317         if (delay <= 0) {
318             if (delay < -per_frame) {
319                 /* printf("grabbing is %d frames late (dropping)\n", (int) -(delay / 16666)); */
320                 s->time_frame += per_frame;
321             }
322             break;
323         }    
324         ts.tv_sec = delay / 1000000;
325         ts.tv_nsec = (delay % 1000000) * 1000;
326         nanosleep(&ts, NULL);
327     }
328
329     if (av_new_packet(pkt, s->frame_size) < 0)
330         return -EIO;
331
332     pkt->pts = curtime & ((1LL << 48) - 1);
333
334     /* read one frame */
335     if (s->aiw_enabled) {
336         return aiw_read_picture(s, pkt->data);
337     } else if (s->use_mmap) {
338         return v4l_mm_read_picture(s, pkt->data);
339     } else {
340         if (read(s->fd, pkt->data, pkt->size) != pkt->size)
341             return -EIO;
342         return s->frame_size;
343     }
344 }
345
346 static int grab_read_close(AVFormatContext *s1)
347 {
348     VideoData *s = s1->priv_data;
349
350     if (s->aiw_enabled)
351         aiw_close(s);
352
353     if (s->use_mmap)
354         munmap(s->video_buf, s->gb_buffers.size);
355
356     /* mute audio. we must force it because the BTTV driver does not
357        return its state correctly */
358     s->audio_saved.flags |= VIDEO_AUDIO_MUTE;
359     ioctl(s->fd, VIDIOCSAUDIO, &s->audio_saved);
360
361     close(s->fd);
362     return 0;
363 }
364
365 static AVInputFormat video_grab_device_format = {
366     "video4linux",
367     "video grab",
368     sizeof(VideoData),
369     NULL,
370     grab_read_header,
371     grab_read_packet,
372     grab_read_close,
373     .flags = AVFMT_NOFILE,
374 };
375
376 /* All in Wonder specific stuff */
377 /* XXX: remove and merge in libavcodec/imgconvert.c */
378
379 static int aiw_init(VideoData *s)
380 {
381     int width, height;
382
383     width = s->width;
384     height = s->height;
385
386     if ((width == s->video_cap.maxwidth && height == s->video_cap.maxheight) ||
387         (width == s->video_cap.maxwidth && height == s->video_cap.maxheight*2) ||
388         (width == s->video_cap.maxwidth/2 && height == s->video_cap.maxheight)) {
389         
390         s->deint=0;
391         s->halfw=0;
392         if (height == s->video_cap.maxheight*2) s->deint=1;
393         if (width == s->video_cap.maxwidth/2) s->halfw=1;
394     } else {
395         av_log(NULL, AV_LOG_ERROR, "\nIncorrect Grab Size Supplied - Supported Sizes Are:\n");
396         av_log(NULL, AV_LOG_ERROR, " %dx%d  %dx%d %dx%d\n\n",
397                 s->video_cap.maxwidth,s->video_cap.maxheight,
398                 s->video_cap.maxwidth,s->video_cap.maxheight*2,
399                 s->video_cap.maxwidth/2,s->video_cap.maxheight);
400         goto fail;
401     }
402
403     if (s->halfw == 0) {
404         s->src_mem = av_malloc(s->width*2);
405     } else {
406         s->src_mem = av_malloc(s->width*4);
407     }
408     if (!s->src_mem) goto fail;
409
410     s->lum_m4_mem = av_malloc(s->width);
411     if (!s->lum_m4_mem)
412         goto fail;
413     return 0;
414  fail:
415     av_freep(&s->src_mem);
416     av_freep(&s->lum_m4_mem);
417     return -1;
418 }
419
420 #ifdef HAVE_MMX
421 #include "../libavcodec/i386/mmx.h"
422
423 #define LINE_WITH_UV \
424                     movq_m2r(ptr[0],mm0); \
425                     movq_m2r(ptr[8],mm1);  \
426                     movq_r2r(mm0, mm4); \
427                     punpcklbw_r2r(mm1,mm0); \
428                     punpckhbw_r2r(mm1,mm4); \
429                     movq_r2r(mm0,mm5); \
430                     punpcklbw_r2r(mm4,mm0); \
431                     punpckhbw_r2r(mm4,mm5); \
432                     movq_r2r(mm0,mm1); \
433                     punpcklbw_r2r(mm5,mm1); \
434                     movq_r2m(mm1,lum[0]); \
435                     movq_m2r(ptr[16],mm2); \
436                     movq_m2r(ptr[24],mm1); \
437                     movq_r2r(mm2,mm4); \
438                     punpcklbw_r2r(mm1,mm2); \
439                     punpckhbw_r2r(mm1,mm4); \
440                     movq_r2r(mm2,mm3); \
441                     punpcklbw_r2r(mm4,mm2); \
442                     punpckhbw_r2r(mm4,mm3); \
443                     movq_r2r(mm2,mm1); \
444                     punpcklbw_r2r(mm3,mm1); \
445                     movq_r2m(mm1,lum[8]); \
446                     punpckhdq_r2r(mm2,mm0); \
447                     punpckhdq_r2r(mm3,mm5); \
448                     movq_r2m(mm0,cb[0]); \
449                     movq_r2m(mm5,cr[0]);
450
451 #define LINE_NO_UV \
452                     movq_m2r(ptr[0],mm0);\
453                     movq_m2r(ptr[8],mm1);\
454                     movq_r2r(mm0, mm4);\
455                     punpcklbw_r2r(mm1,mm0); \
456                     punpckhbw_r2r(mm1,mm4);\
457                     movq_r2r(mm0,mm5);\
458                     punpcklbw_r2r(mm4,mm0);\
459                     punpckhbw_r2r(mm4,mm5);\
460                     movq_r2r(mm0,mm1);\
461                     punpcklbw_r2r(mm5,mm1);\
462                     movq_r2m(mm1,lum[0]);\
463                     movq_m2r(ptr[16],mm2);\
464                     movq_m2r(ptr[24],mm1);\
465                     movq_r2r(mm2,mm4);\
466                     punpcklbw_r2r(mm1,mm2);\
467                     punpckhbw_r2r(mm1,mm4);\
468                     movq_r2r(mm2,mm3);\
469                     punpcklbw_r2r(mm4,mm2);\
470                     punpckhbw_r2r(mm4,mm3);\
471                     movq_r2r(mm2,mm1);\
472                     punpcklbw_r2r(mm3,mm1);\
473                     movq_r2m(mm1,lum[8]);
474
475 #define LINE_WITHUV_AVG \
476                     movq_m2r(ptr[0], mm0);\
477                     movq_m2r(ptr[8], mm1);\
478                     movq_r2r(mm0, mm4);\
479                     punpcklbw_r2r(mm1,mm0);\
480                     punpckhbw_r2r(mm1,mm4);\
481                     movq_r2r(mm0,mm5);\
482                     punpcklbw_r2r(mm4,mm0);\
483                     punpckhbw_r2r(mm4,mm5);\
484                     movq_r2r(mm0,mm1);\
485                     movq_r2r(mm5,mm2);\
486                     punpcklbw_r2r(mm7,mm1);\
487                     punpcklbw_r2r(mm7,mm2);\
488                     paddw_r2r(mm6,mm1);\
489                     paddw_r2r(mm2,mm1);\
490                     psraw_i2r(1,mm1);\
491                     packuswb_r2r(mm7,mm1);\
492                     movd_r2m(mm1,lum[0]);\
493                     movq_m2r(ptr[16],mm2);\
494                     movq_m2r(ptr[24],mm1);\
495                     movq_r2r(mm2,mm4);\
496                     punpcklbw_r2r(mm1,mm2);\
497                     punpckhbw_r2r(mm1,mm4);\
498                     movq_r2r(mm2,mm3);\
499                     punpcklbw_r2r(mm4,mm2);\
500                     punpckhbw_r2r(mm4,mm3);\
501                     movq_r2r(mm2,mm1);\
502                     movq_r2r(mm3,mm4);\
503                     punpcklbw_r2r(mm7,mm1);\
504                     punpcklbw_r2r(mm7,mm4);\
505                     paddw_r2r(mm6,mm1);\
506                     paddw_r2r(mm4,mm1);\
507                     psraw_i2r(1,mm1);\
508                     packuswb_r2r(mm7,mm1);\
509                     movd_r2m(mm1,lum[4]);\
510                     punpckhbw_r2r(mm7,mm0);\
511                     punpckhbw_r2r(mm7,mm2);\
512                     paddw_r2r(mm6,mm0);\
513                     paddw_r2r(mm2,mm0);\
514                     psraw_i2r(1,mm0);\
515                     packuswb_r2r(mm7,mm0);\
516                     punpckhbw_r2r(mm7,mm5);\
517                     punpckhbw_r2r(mm7,mm3);\
518                     paddw_r2r(mm6,mm5);\
519                     paddw_r2r(mm3,mm5);\
520                     psraw_i2r(1,mm5);\
521                     packuswb_r2r(mm7,mm5);\
522                     movd_r2m(mm0,cb[0]);\
523                     movd_r2m(mm5,cr[0]);
524
525 #define LINE_NOUV_AVG \
526                     movq_m2r(ptr[0],mm0);\
527                     movq_m2r(ptr[8],mm1);\
528                     pand_r2r(mm5,mm0);\
529                     pand_r2r(mm5,mm1);\
530                     pmaddwd_r2r(mm6,mm0);\
531                     pmaddwd_r2r(mm6,mm1);\
532                     packssdw_r2r(mm1,mm0);\
533                     paddw_r2r(mm6,mm0);\
534                     psraw_i2r(1,mm0);\
535                     movq_m2r(ptr[16],mm2);\
536                     movq_m2r(ptr[24],mm3);\
537                     pand_r2r(mm5,mm2);\
538                     pand_r2r(mm5,mm3);\
539                     pmaddwd_r2r(mm6,mm2);\
540                     pmaddwd_r2r(mm6,mm3);\
541                     packssdw_r2r(mm3,mm2);\
542                     paddw_r2r(mm6,mm2);\
543                     psraw_i2r(1,mm2);\
544                     packuswb_r2r(mm2,mm0);\
545                     movq_r2m(mm0,lum[0]);
546
547 #define DEINT_LINE_LUM(ptroff) \
548                     movd_m2r(lum_m4[(ptroff)],mm0);\
549                     movd_m2r(lum_m3[(ptroff)],mm1);\
550                     movd_m2r(lum_m2[(ptroff)],mm2);\
551                     movd_m2r(lum_m1[(ptroff)],mm3);\
552                     movd_m2r(lum[(ptroff)],mm4);\
553                     punpcklbw_r2r(mm7,mm0);\
554                     movd_r2m(mm2,lum_m4[(ptroff)]);\
555                     punpcklbw_r2r(mm7,mm1);\
556                     punpcklbw_r2r(mm7,mm2);\
557                     punpcklbw_r2r(mm7,mm3);\
558                     punpcklbw_r2r(mm7,mm4);\
559                     psllw_i2r(2,mm1);\
560                     psllw_i2r(1,mm2);\
561                     paddw_r2r(mm6,mm1);\
562                     psllw_i2r(2,mm3);\
563                     paddw_r2r(mm2,mm1);\
564                     paddw_r2r(mm4,mm0);\
565                     paddw_r2r(mm3,mm1);\
566                     psubusw_r2r(mm0,mm1);\
567                     psrlw_i2r(3,mm1);\
568                     packuswb_r2r(mm7,mm1);\
569                     movd_r2m(mm1,lum_m2[(ptroff)]);
570
571 #else
572 #include "../libavcodec/dsputil.h"
573
574 #define LINE_WITH_UV \
575                     lum[0]=ptr[0];lum[1]=ptr[2];lum[2]=ptr[4];lum[3]=ptr[6];\
576                     cb[0]=ptr[1];cb[1]=ptr[5];\
577                     cr[0]=ptr[3];cr[1]=ptr[7];\
578                     lum[4]=ptr[8];lum[5]=ptr[10];lum[6]=ptr[12];lum[7]=ptr[14];\
579                     cb[2]=ptr[9];cb[3]=ptr[13];\
580                     cr[2]=ptr[11];cr[3]=ptr[15];\
581                     lum[8]=ptr[16];lum[9]=ptr[18];lum[10]=ptr[20];lum[11]=ptr[22];\
582                     cb[4]=ptr[17];cb[5]=ptr[21];\
583                     cr[4]=ptr[19];cr[5]=ptr[23];\
584                     lum[12]=ptr[24];lum[13]=ptr[26];lum[14]=ptr[28];lum[15]=ptr[30];\
585                     cb[6]=ptr[25];cb[7]=ptr[29];\
586                     cr[6]=ptr[27];cr[7]=ptr[31];
587
588 #define LINE_NO_UV \
589                     lum[0]=ptr[0];lum[1]=ptr[2];lum[2]=ptr[4];lum[3]=ptr[6];\
590                     lum[4]=ptr[8];lum[5]=ptr[10];lum[6]=ptr[12];lum[7]=ptr[14];\
591                     lum[8]=ptr[16];lum[9]=ptr[18];lum[10]=ptr[20];lum[11]=ptr[22];\
592                     lum[12]=ptr[24];lum[13]=ptr[26];lum[14]=ptr[28];lum[15]=ptr[30];
593
594 #define LINE_WITHUV_AVG \
595                     sum=(ptr[0]+ptr[2]+1) >> 1;lum[0]=sum; \
596                     sum=(ptr[4]+ptr[6]+1) >> 1;lum[1]=sum; \
597                     sum=(ptr[1]+ptr[5]+1) >> 1;cb[0]=sum; \
598                     sum=(ptr[3]+ptr[7]+1) >> 1;cr[0]=sum; \
599                     sum=(ptr[8]+ptr[10]+1) >> 1;lum[2]=sum; \
600                     sum=(ptr[12]+ptr[14]+1) >> 1;lum[3]=sum; \
601                     sum=(ptr[9]+ptr[13]+1) >> 1;cb[1]=sum; \
602                     sum=(ptr[11]+ptr[15]+1) >> 1;cr[1]=sum; \
603                     sum=(ptr[16]+ptr[18]+1) >> 1;lum[4]=sum; \
604                     sum=(ptr[20]+ptr[22]+1) >> 1;lum[5]=sum; \
605                     sum=(ptr[17]+ptr[21]+1) >> 1;cb[2]=sum; \
606                     sum=(ptr[19]+ptr[23]+1) >> 1;cr[2]=sum; \
607                     sum=(ptr[24]+ptr[26]+1) >> 1;lum[6]=sum; \
608                     sum=(ptr[28]+ptr[30]+1) >> 1;lum[7]=sum; \
609                     sum=(ptr[25]+ptr[29]+1) >> 1;cb[3]=sum; \
610                     sum=(ptr[27]+ptr[31]+1) >> 1;cr[3]=sum; 
611
612 #define LINE_NOUV_AVG \
613                     sum=(ptr[0]+ptr[2]+1) >> 1;lum[0]=sum; \
614                     sum=(ptr[4]+ptr[6]+1) >> 1;lum[1]=sum; \
615                     sum=(ptr[8]+ptr[10]+1) >> 1;lum[2]=sum; \
616                     sum=(ptr[12]+ptr[14]+1) >> 1;lum[3]=sum; \
617                     sum=(ptr[16]+ptr[18]+1) >> 1;lum[4]=sum; \
618                     sum=(ptr[20]+ptr[22]+1) >> 1;lum[5]=sum; \
619                     sum=(ptr[24]+ptr[26]+1) >> 1;lum[6]=sum; \
620                     sum=(ptr[28]+ptr[30]+1) >> 1;lum[7]=sum; 
621
622 #define DEINT_LINE_LUM(ptroff) \
623                     sum=(-lum_m4[(ptroff)]+(lum_m3[(ptroff)]<<2)+(lum_m2[(ptroff)]<<1)+(lum_m1[(ptroff)]<<2)-lum[(ptroff)]); \
624                     lum_m4[(ptroff)]=lum_m2[(ptroff)];\
625                     lum_m2[(ptroff)]=cm[(sum+4)>>3];\
626                     sum=(-lum_m4[(ptroff)+1]+(lum_m3[(ptroff)+1]<<2)+(lum_m2[(ptroff)+1]<<1)+(lum_m1[(ptroff)+1]<<2)-lum[(ptroff)+1]); \
627                     lum_m4[(ptroff)+1]=lum_m2[(ptroff)+1];\
628                     lum_m2[(ptroff)+1]=cm[(sum+4)>>3];\
629                     sum=(-lum_m4[(ptroff)+2]+(lum_m3[(ptroff)+2]<<2)+(lum_m2[(ptroff)+2]<<1)+(lum_m1[(ptroff)+2]<<2)-lum[(ptroff)+2]); \
630                     lum_m4[(ptroff)+2]=lum_m2[(ptroff)+2];\
631                     lum_m2[(ptroff)+2]=cm[(sum+4)>>3];\
632                     sum=(-lum_m4[(ptroff)+3]+(lum_m3[(ptroff)+3]<<2)+(lum_m2[(ptroff)+3]<<1)+(lum_m1[(ptroff)+3]<<2)-lum[(ptroff)+3]); \
633                     lum_m4[(ptroff)+3]=lum_m2[(ptroff)+3];\
634                     lum_m2[(ptroff)+3]=cm[(sum+4)>>3];
635
636 #endif
637
638
639 /* Read two fields separately. */
640 static int aiw_read_picture(VideoData *s, uint8_t *data)
641 {
642     uint8_t *ptr, *lum, *cb, *cr;
643     int h;
644 #ifndef HAVE_MMX
645     int sum;
646 #endif
647     uint8_t* src = s->src_mem;
648     uint8_t *ptrend = &src[s->width*2];
649     lum=data;
650     cb=&lum[s->width*s->height];
651     cr=&cb[(s->width*s->height)/4];
652     if (s->deint == 0 && s->halfw == 0) {
653         while (read(s->fd,src,s->width*2) < 0) {
654             usleep(100);
655         }
656         for (h = 0; h < s->height-2; h+=2) {
657             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
658                 LINE_WITH_UV
659                     }
660             read(s->fd,src,s->width*2);
661             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16) {
662                 LINE_NO_UV
663                     }
664             read(s->fd,src,s->width*2);
665         }
666         /*
667          * Do last two lines
668          */
669         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
670             LINE_WITH_UV
671                 }
672         read(s->fd,src,s->width*2);
673         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16) {
674             LINE_NO_UV
675                 }
676         /* drop second field */
677         while (read(s->fd,src,s->width*2) < 0) {
678             usleep(100);
679         }
680         for (h = 0; h < s->height - 1; h++) {
681             read(s->fd,src,s->width*2);
682         }
683     } else if (s->halfw == 1) {
684 #ifdef HAVE_MMX
685         mmx_t rounder;
686         mmx_t masker;
687         rounder.uw[0]=1;
688         rounder.uw[1]=1;
689         rounder.uw[2]=1;
690         rounder.uw[3]=1;
691         masker.ub[0]=0xff;
692         masker.ub[1]=0;
693         masker.ub[2]=0xff;
694         masker.ub[3]=0;
695         masker.ub[4]=0xff;
696         masker.ub[5]=0;
697         masker.ub[6]=0xff;
698         masker.ub[7]=0;
699         pxor_r2r(mm7,mm7);
700         movq_m2r(rounder,mm6);
701 #endif
702         while (read(s->fd,src,s->width*4) < 0) {
703             usleep(100);
704         }
705         ptrend = &src[s->width*4];
706         for (h = 0; h < s->height-2; h+=2) {
707             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=8, cb+=4, cr+=4) {
708                 LINE_WITHUV_AVG
709                     }
710             read(s->fd,src,s->width*4);
711 #ifdef HAVE_MMX
712             movq_m2r(masker,mm5);
713 #endif
714             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=8) {
715                 LINE_NOUV_AVG
716                     }
717             read(s->fd,src,s->width*4);
718         }
719         /*
720  * Do last two lines
721  */
722         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=8, cb+=4, cr+=4) {
723             LINE_WITHUV_AVG
724                 }
725         read(s->fd,src,s->width*4);
726 #ifdef HAVE_MMX
727         movq_m2r(masker,mm5);
728 #endif
729         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=8) {
730             LINE_NOUV_AVG
731                 }
732         /* drop second field */
733         while (read(s->fd,src,s->width*4) < 0) {
734             usleep(100);
735         }
736         for (h = 0; h < s->height - 1; h++) {
737             read(s->fd,src,s->width*4);
738         }
739     } else {
740         uint8_t *lum_m1, *lum_m2, *lum_m3, *lum_m4;
741 #ifdef HAVE_MMX
742         mmx_t rounder;
743         rounder.uw[0]=4;
744         rounder.uw[1]=4;
745         rounder.uw[2]=4;
746         rounder.uw[3]=4;
747         movq_m2r(rounder,mm6);
748         pxor_r2r(mm7,mm7);
749 #else
750         uint8_t *cm = cropTbl + MAX_NEG_CROP;
751 #endif
752
753         /* read two fields and deinterlace them */
754         while (read(s->fd,src,s->width*2) < 0) {
755             usleep(100);
756         }
757         for (h = 0; h < (s->height/2)-2; h+=2) {
758             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
759                 LINE_WITH_UV
760                     }
761             read(s->fd,src,s->width*2);
762             /* skip a luminance line - will be filled in later */
763             lum += s->width;
764             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
765                 LINE_WITH_UV
766                     }
767             /* skip a luminance line - will be filled in later */
768             lum += s->width;
769             read(s->fd,src,s->width*2);
770         }
771         /*
772  * Do last two lines
773  */
774         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
775             LINE_WITH_UV
776                 }
777         /* skip a luminance line - will be filled in later */
778         lum += s->width;
779         read(s->fd,src,s->width*2);
780         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
781             LINE_WITH_UV
782                 }
783         /*
784  *
785  * SECOND FIELD
786  *
787  */
788         lum=&data[s->width];
789         while (read(s->fd,src,s->width*2) < 0) {
790             usleep(10);
791         }
792         /* First (and last) two lines not interlaced */
793         for (h = 0; h < 2; h++) {
794             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16) {
795                 LINE_NO_UV
796                     }
797             read(s->fd,src,s->width*2);
798             /* skip a luminance line */
799             lum += s->width;
800         }
801         lum_m1=&lum[-s->width];
802         lum_m2=&lum_m1[-s->width];
803         lum_m3=&lum_m2[-s->width];
804         memmove(s->lum_m4_mem,&lum_m3[-s->width],s->width);
805         for (; h < (s->height/2)-1; h++) {
806             lum_m4=s->lum_m4_mem;
807             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16,lum_m1+=16,lum_m2+=16,lum_m3+=16,lum_m4+=16) {
808                 LINE_NO_UV
809
810                     DEINT_LINE_LUM(0)
811                     DEINT_LINE_LUM(4)
812                     DEINT_LINE_LUM(8)
813                     DEINT_LINE_LUM(12)
814                     }
815             read(s->fd,src,s->width*2);
816             /* skip a luminance line */
817             lum += s->width;
818             lum_m1 += s->width;
819             lum_m2 += s->width;
820             lum_m3 += s->width;
821             //                lum_m4 += s->width;
822         }
823         /*
824  * Do last line
825  */
826         lum_m4=s->lum_m4_mem;
827         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, lum_m1+=16, lum_m2+=16, lum_m3+=16, lum_m4+=16) {
828             LINE_NO_UV
829
830                 DEINT_LINE_LUM(0)
831                 DEINT_LINE_LUM(4)
832                 DEINT_LINE_LUM(8)
833                 DEINT_LINE_LUM(12)
834                 }
835     }
836 #ifdef HAVE_MMX
837     emms();
838 #endif
839     return s->frame_size;
840 }
841
842 static int aiw_close(VideoData *s)
843 {
844     av_freep(&s->lum_m4_mem);
845     av_freep(&s->src_mem);
846     return 0;
847 }
848
849 int video_grab_init(void)
850 {
851     av_register_input_format(&video_grab_device_format);
852     return 0;
853 }