git.sesse.net Git - ffmpeg/blob - libavformat/grab.c

   1 /*
   2  * Linux video grab interface
   3  * Copyright (c) 2000,2001 Fabrice Bellard.
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21 #include "avformat.h"
  22 #include <unistd.h>
  23 #include <fcntl.h>
  24 #include <sys/ioctl.h>
  25 #include <sys/mman.h>
  26 #include <sys/time.h>
  27 #define _LINUX_TIME_H 1
  28 #include <linux/videodev.h>
  29 #include <time.h>
  30
  31 typedef struct {
  32     int fd;
  33     int frame_format; /* see VIDEO_PALETTE_xxx */
  34     int use_mmap;
  35     int width, height;
  36     int frame_rate;
  37     int frame_rate_base;
  38     int64_t time_frame;
  39     int frame_size;
  40     struct video_capability video_cap;
  41     struct video_audio audio_saved;
  42     uint8_t *video_buf;
  43     struct video_mbuf gb_buffers;
  44     struct video_mmap gb_buf;
  45     int gb_frame;
  46
  47     /* ATI All In Wonder specific stuff */
  48     /* XXX: remove and merge in libavcodec/imgconvert.c */
  49     int aiw_enabled;
  50     int deint;
  51     int halfw;
  52     uint8_t *src_mem;
  53     uint8_t *lum_m4_mem;
  54 } VideoData;
  55
  56 static int aiw_init(VideoData *s);
  57 static int aiw_read_picture(VideoData *s, uint8_t *data);
  58 static int aiw_close(VideoData *s);
  59
  60 static int grab_read_header(AVFormatContext *s1, AVFormatParameters *ap)
  61 {
  62     VideoData *s = s1->priv_data;
  63     AVStream *st;
  64     int width, height;
  65     int video_fd, frame_size;
  66     int ret, frame_rate, frame_rate_base;
  67     int desired_palette, desired_depth;
  68     struct video_tuner tuner;
  69     struct video_audio audio;
  70     struct video_picture pict;
  71     int j;
  72
  73     if (ap->width <= 0 || ap->height <= 0 || ap->time_base.den <= 0) {
  74         av_log(s1, AV_LOG_ERROR, "Bad capture size (%dx%d) or wrong time base (%d)\n",
  75             ap->width, ap->height, ap->time_base.den);
  76
  77         return -1;
  78     }
  79
  80     width = ap->width;
  81     height = ap->height;
  82     frame_rate      = ap->time_base.den;
  83     frame_rate_base = ap->time_base.num;
  84
  85     if((unsigned)width > 32767 || (unsigned)height > 32767) {
  86         av_log(s1, AV_LOG_ERROR, "Capture size is out of range: %dx%d\n",
  87             width, height);
  88
  89         return -1;
  90     }
  91
  92     st = av_new_stream(s1, 0);
  93     if (!st)
  94         return AVERROR(ENOMEM);
  95     av_set_pts_info(st, 64, 1, 1000000); /* 64 bits pts in us */
  96
  97     s->width = width;
  98     s->height = height;
  99     s->frame_rate      = frame_rate;
 100     s->frame_rate_base = frame_rate_base;
 101
 102     video_fd = open(s1->filename, O_RDWR);
 103     if (video_fd < 0) {
 104         perror(s1->filename);
 105         goto fail;
 106     }
 107
 108     if (ioctl(video_fd,VIDIOCGCAP, &s->video_cap) < 0) {
 109         perror("VIDIOCGCAP");
 110         goto fail;
 111     }
 112
 113     if (!(s->video_cap.type & VID_TYPE_CAPTURE)) {
 114         av_log(s1, AV_LOG_ERROR, "Fatal: grab device does not handle capture\n");
 115         goto fail;
 116     }
 117
 118     desired_palette = -1;
 119     desired_depth = -1;
 120     if (ap->pix_fmt == PIX_FMT_YUV420P) {
 121         desired_palette = VIDEO_PALETTE_YUV420P;
 122         desired_depth = 12;
 123     } else if (ap->pix_fmt == PIX_FMT_YUYV422) {
 124         desired_palette = VIDEO_PALETTE_YUV422;
 125         desired_depth = 16;
 126     } else if (ap->pix_fmt == PIX_FMT_BGR24) {
 127         desired_palette = VIDEO_PALETTE_RGB24;
 128         desired_depth = 24;
 129     }
 130
 131     /* set tv standard */
 132     if (ap->standard && !ioctl(video_fd, VIDIOCGTUNER, &tuner)) {
 133         if (!strcasecmp(ap->standard, "pal"))
 134             tuner.mode = VIDEO_MODE_PAL;
 135         else if (!strcasecmp(ap->standard, "secam"))
 136             tuner.mode = VIDEO_MODE_SECAM;
 137         else
 138             tuner.mode = VIDEO_MODE_NTSC;
 139         ioctl(video_fd, VIDIOCSTUNER, &tuner);
 140     }
 141
 142     /* unmute audio */
 143     audio.audio = 0;
 144     ioctl(video_fd, VIDIOCGAUDIO, &audio);
 145     memcpy(&s->audio_saved, &audio, sizeof(audio));
 146     audio.flags &= ~VIDEO_AUDIO_MUTE;
 147     ioctl(video_fd, VIDIOCSAUDIO, &audio);
 148
 149     ioctl(video_fd, VIDIOCGPICT, &pict);
 150 #if 0
 151     printf("v4l: colour=%d hue=%d brightness=%d constrast=%d whiteness=%d\n",
 152            pict.colour,
 153            pict.hue,
 154            pict.brightness,
 155            pict.contrast,
 156            pict.whiteness);
 157 #endif
 158     /* try to choose a suitable video format */
 159     pict.palette = desired_palette;
 160     pict.depth= desired_depth;
 161     if (desired_palette == -1 || (ret = ioctl(video_fd, VIDIOCSPICT, &pict)) < 0) {
 162         pict.palette=VIDEO_PALETTE_YUV420P;
 163         pict.depth=12;
 164         ret = ioctl(video_fd, VIDIOCSPICT, &pict);
 165         if (ret < 0) {
 166             pict.palette=VIDEO_PALETTE_YUV422;
 167             pict.depth=16;
 168             ret = ioctl(video_fd, VIDIOCSPICT, &pict);
 169             if (ret < 0) {
 170                 pict.palette=VIDEO_PALETTE_RGB24;
 171                 pict.depth=24;
 172                 ret = ioctl(video_fd, VIDIOCSPICT, &pict);
 173                 if (ret < 0) {
 174                     pict.palette=VIDEO_PALETTE_GREY;
 175                     pict.depth=8;
 176                     ret = ioctl(video_fd, VIDIOCSPICT, &pict);
 177                     if (ret < 0)
 178                         goto fail1;
 179                 }
 180             }
 181         }
 182     }
 183
 184     ret = ioctl(video_fd,VIDIOCGMBUF,&s->gb_buffers);
 185     if (ret < 0) {
 186         /* try to use read based access */
 187         struct video_window win;
 188         int val;
 189
 190         win.x = 0;
 191         win.y = 0;
 192         win.width = width;
 193         win.height = height;
 194         win.chromakey = -1;
 195         win.flags = 0;
 196
 197         ioctl(video_fd, VIDIOCSWIN, &win);
 198
 199         s->frame_format = pict.palette;
 200
 201         val = 1;
 202         ioctl(video_fd, VIDIOCCAPTURE, &val);
 203
 204         s->time_frame = av_gettime() * s->frame_rate / s->frame_rate_base;
 205         s->use_mmap = 0;
 206
 207         /* ATI All In Wonder automatic activation */
 208         if (!strcmp(s->video_cap.name, "Km")) {
 209             if (aiw_init(s) < 0)
 210                 goto fail;
 211             s->aiw_enabled = 1;
 212             /* force 420P format because convertion from YUV422 to YUV420P
 213                is done in this driver (ugly) */
 214             s->frame_format = VIDEO_PALETTE_YUV420P;
 215         }
 216     } else {
 217         s->video_buf = mmap(0,s->gb_buffers.size,PROT_READ|PROT_WRITE,MAP_SHARED,video_fd,0);
 218         if ((unsigned char*)-1 == s->video_buf) {
 219             s->video_buf = mmap(0,s->gb_buffers.size,PROT_READ|PROT_WRITE,MAP_PRIVATE,video_fd,0);
 220             if ((unsigned char*)-1 == s->video_buf) {
 221                 perror("mmap");
 222                 goto fail;
 223             }
 224         }
 225         s->gb_frame = 0;
 226         s->time_frame = av_gettime() * s->frame_rate / s->frame_rate_base;
 227
 228         /* start to grab the first frame */
 229         s->gb_buf.frame = s->gb_frame % s->gb_buffers.frames;
 230         s->gb_buf.height = height;
 231         s->gb_buf.width = width;
 232         s->gb_buf.format = pict.palette;
 233
 234         ret = ioctl(video_fd, VIDIOCMCAPTURE, &s->gb_buf);
 235         if (ret < 0) {
 236             if (errno != EAGAIN) {
 237             fail1:
 238                 av_log(s1, AV_LOG_ERROR, "Fatal: grab device does not support suitable format\n");
 239             } else {
 240                 av_log(s1, AV_LOG_ERROR,"Fatal: grab device does not receive any video signal\n");
 241             }
 242             goto fail;
 243         }
 244         for (j = 1; j < s->gb_buffers.frames; j++) {
 245           s->gb_buf.frame = j;
 246           ioctl(video_fd, VIDIOCMCAPTURE, &s->gb_buf);
 247         }
 248         s->frame_format = s->gb_buf.format;
 249         s->use_mmap = 1;
 250     }
 251
 252     switch(s->frame_format) {
 253     case VIDEO_PALETTE_YUV420P:
 254         frame_size = (width * height * 3) / 2;
 255         st->codec->pix_fmt = PIX_FMT_YUV420P;
 256         break;
 257     case VIDEO_PALETTE_YUV422:
 258         frame_size = width * height * 2;
 259         st->codec->pix_fmt = PIX_FMT_YUYV422;
 260         break;
 261     case VIDEO_PALETTE_RGB24:
 262         frame_size = width * height * 3;
 263         st->codec->pix_fmt = PIX_FMT_BGR24; /* NOTE: v4l uses BGR24, not RGB24 ! */
 264         break;
 265     case VIDEO_PALETTE_GREY:
 266         frame_size = width * height * 1;
 267         st->codec->pix_fmt = PIX_FMT_GRAY8;
 268         break;
 269     default:
 270         goto fail;
 271     }
 272     s->fd = video_fd;
 273     s->frame_size = frame_size;
 274
 275     st->codec->codec_type = CODEC_TYPE_VIDEO;
 276     st->codec->codec_id = CODEC_ID_RAWVIDEO;
 277     st->codec->width = width;
 278     st->codec->height = height;
 279     st->codec->time_base.den      = frame_rate;
 280     st->codec->time_base.num = frame_rate_base;
 281     st->codec->bit_rate = frame_size * 1/av_q2d(st->codec->time_base) * 8;
 282
 283     return 0;
 284  fail:
 285     if (video_fd >= 0)
 286         close(video_fd);
 287     av_free(st);
 288     return AVERROR_IO;
 289 }
 290
 291 static int v4l_mm_read_picture(VideoData *s, uint8_t *buf)
 292 {
 293     uint8_t *ptr;
 294
 295     while (ioctl(s->fd, VIDIOCSYNC, &s->gb_frame) < 0 &&
 296            (errno == EAGAIN || errno == EINTR));
 297
 298     ptr = s->video_buf + s->gb_buffers.offsets[s->gb_frame];
 299     memcpy(buf, ptr, s->frame_size);
 300
 301     /* Setup to capture the next frame */
 302     s->gb_buf.frame = s->gb_frame;
 303     if (ioctl(s->fd, VIDIOCMCAPTURE, &s->gb_buf) < 0) {
 304         if (errno == EAGAIN)
 305             av_log(NULL, AV_LOG_ERROR, "Cannot Sync\n");
 306         else
 307             perror("VIDIOCMCAPTURE");
 308         return AVERROR_IO;
 309     }
 310
 311     /* This is now the grabbing frame */
 312     s->gb_frame = (s->gb_frame + 1) % s->gb_buffers.frames;
 313
 314     return s->frame_size;
 315 }
 316
 317 static int grab_read_packet(AVFormatContext *s1, AVPacket *pkt)
 318 {
 319     VideoData *s = s1->priv_data;
 320     int64_t curtime, delay;
 321     struct timespec ts;
 322
 323     /* Calculate the time of the next frame */
 324     s->time_frame += INT64_C(1000000);
 325
 326     /* wait based on the frame rate */
 327     for(;;) {
 328         curtime = av_gettime();
 329         delay = s->time_frame  * s->frame_rate_base / s->frame_rate - curtime;
 330         if (delay <= 0) {
 331             if (delay < INT64_C(-1000000) * s->frame_rate_base / s->frame_rate) {
 332                 /* printf("grabbing is %d frames late (dropping)\n", (int) -(delay / 16666)); */
 333                 s->time_frame += INT64_C(1000000);
 334             }
 335             break;
 336         }
 337         ts.tv_sec = delay / 1000000;
 338         ts.tv_nsec = (delay % 1000000) * 1000;
 339         nanosleep(&ts, NULL);
 340     }
 341
 342     if (av_new_packet(pkt, s->frame_size) < 0)
 343         return AVERROR_IO;
 344
 345     pkt->pts = curtime;
 346
 347     /* read one frame */
 348     if (s->aiw_enabled) {
 349         return aiw_read_picture(s, pkt->data);
 350     } else if (s->use_mmap) {
 351         return v4l_mm_read_picture(s, pkt->data);
 352     } else {
 353         if (read(s->fd, pkt->data, pkt->size) != pkt->size)
 354             return AVERROR_IO;
 355         return s->frame_size;
 356     }
 357 }
 358
 359 static int grab_read_close(AVFormatContext *s1)
 360 {
 361     VideoData *s = s1->priv_data;
 362
 363     if (s->aiw_enabled)
 364         aiw_close(s);
 365
 366     if (s->use_mmap)
 367         munmap(s->video_buf, s->gb_buffers.size);
 368
 369     /* mute audio. we must force it because the BTTV driver does not
 370        return its state correctly */
 371     s->audio_saved.flags |= VIDEO_AUDIO_MUTE;
 372     ioctl(s->fd, VIDIOCSAUDIO, &s->audio_saved);
 373
 374     close(s->fd);
 375     return 0;
 376 }
 377
 378 AVInputFormat video_grab_v4l_demuxer = {
 379     "video4linux",
 380     "video grab",
 381     sizeof(VideoData),
 382     NULL,
 383     grab_read_header,
 384     grab_read_packet,
 385     grab_read_close,
 386     .flags = AVFMT_NOFILE,
 387 };
 388
 389 /* All in Wonder specific stuff */
 390 /* XXX: remove and merge in libavcodec/imgconvert.c */
 391
 392 static int aiw_init(VideoData *s)
 393 {
 394     int width, height;
 395
 396     width = s->width;
 397     height = s->height;
 398
 399     if ((width == s->video_cap.maxwidth && height == s->video_cap.maxheight) ||
 400         (width == s->video_cap.maxwidth && height == s->video_cap.maxheight*2) ||
 401         (width == s->video_cap.maxwidth/2 && height == s->video_cap.maxheight)) {
 402
 403         s->deint=0;
 404         s->halfw=0;
 405         if (height == s->video_cap.maxheight*2) s->deint=1;
 406         if (width == s->video_cap.maxwidth/2) s->halfw=1;
 407     } else {
 408         av_log(NULL, AV_LOG_ERROR, "\nIncorrect Grab Size Supplied - Supported Sizes Are:\n");
 409         av_log(NULL, AV_LOG_ERROR, " %dx%d  %dx%d %dx%d\n\n",
 410                 s->video_cap.maxwidth,s->video_cap.maxheight,
 411                 s->video_cap.maxwidth,s->video_cap.maxheight*2,
 412                 s->video_cap.maxwidth/2,s->video_cap.maxheight);
 413         goto fail;
 414     }
 415
 416     if (s->halfw == 0) {
 417         s->src_mem = av_malloc(s->width*2);
 418     } else {
 419         s->src_mem = av_malloc(s->width*4);
 420     }
 421     if (!s->src_mem) goto fail;
 422
 423     s->lum_m4_mem = av_malloc(s->width);
 424     if (!s->lum_m4_mem)
 425         goto fail;
 426     return 0;
 427  fail:
 428     av_freep(&s->src_mem);
 429     av_freep(&s->lum_m4_mem);
 430     return -1;
 431 }
 432
 433 #ifdef HAVE_MMX
 434 #include "i386/mmx.h"
 435
 436 #define LINE_WITH_UV \
 437                     movq_m2r(ptr[0],mm0); \
 438                     movq_m2r(ptr[8],mm1);  \
 439                     movq_r2r(mm0, mm4); \
 440                     punpcklbw_r2r(mm1,mm0); \
 441                     punpckhbw_r2r(mm1,mm4); \
 442                     movq_r2r(mm0,mm5); \
 443                     punpcklbw_r2r(mm4,mm0); \
 444                     punpckhbw_r2r(mm4,mm5); \
 445                     movq_r2r(mm0,mm1); \
 446                     punpcklbw_r2r(mm5,mm1); \
 447                     movq_r2m(mm1,lum[0]); \
 448                     movq_m2r(ptr[16],mm2); \
 449                     movq_m2r(ptr[24],mm1); \
 450                     movq_r2r(mm2,mm4); \
 451                     punpcklbw_r2r(mm1,mm2); \
 452                     punpckhbw_r2r(mm1,mm4); \
 453                     movq_r2r(mm2,mm3); \
 454                     punpcklbw_r2r(mm4,mm2); \
 455                     punpckhbw_r2r(mm4,mm3); \
 456                     movq_r2r(mm2,mm1); \
 457                     punpcklbw_r2r(mm3,mm1); \
 458                     movq_r2m(mm1,lum[8]); \
 459                     punpckhdq_r2r(mm2,mm0); \
 460                     punpckhdq_r2r(mm3,mm5); \
 461                     movq_r2m(mm0,cb[0]); \
 462                     movq_r2m(mm5,cr[0]);
 463
 464 #define LINE_NO_UV \
 465                     movq_m2r(ptr[0],mm0);\
 466                     movq_m2r(ptr[8],mm1);\
 467                     movq_r2r(mm0, mm4);\
 468                     punpcklbw_r2r(mm1,mm0); \
 469                     punpckhbw_r2r(mm1,mm4);\
 470                     movq_r2r(mm0,mm5);\
 471                     punpcklbw_r2r(mm4,mm0);\
 472                     punpckhbw_r2r(mm4,mm5);\
 473                     movq_r2r(mm0,mm1);\
 474                     punpcklbw_r2r(mm5,mm1);\
 475                     movq_r2m(mm1,lum[0]);\
 476                     movq_m2r(ptr[16],mm2);\
 477                     movq_m2r(ptr[24],mm1);\
 478                     movq_r2r(mm2,mm4);\
 479                     punpcklbw_r2r(mm1,mm2);\
 480                     punpckhbw_r2r(mm1,mm4);\
 481                     movq_r2r(mm2,mm3);\
 482                     punpcklbw_r2r(mm4,mm2);\
 483                     punpckhbw_r2r(mm4,mm3);\
 484                     movq_r2r(mm2,mm1);\
 485                     punpcklbw_r2r(mm3,mm1);\
 486                     movq_r2m(mm1,lum[8]);
 487
 488 #define LINE_WITHUV_AVG \
 489                     movq_m2r(ptr[0], mm0);\
 490                     movq_m2r(ptr[8], mm1);\
 491                     movq_r2r(mm0, mm4);\
 492                     punpcklbw_r2r(mm1,mm0);\
 493                     punpckhbw_r2r(mm1,mm4);\
 494                     movq_r2r(mm0,mm5);\
 495                     punpcklbw_r2r(mm4,mm0);\
 496                     punpckhbw_r2r(mm4,mm5);\
 497                     movq_r2r(mm0,mm1);\
 498                     movq_r2r(mm5,mm2);\
 499                     punpcklbw_r2r(mm7,mm1);\
 500                     punpcklbw_r2r(mm7,mm2);\
 501                     paddw_r2r(mm6,mm1);\
 502                     paddw_r2r(mm2,mm1);\
 503                     psraw_i2r(1,mm1);\
 504                     packuswb_r2r(mm7,mm1);\
 505                     movd_r2m(mm1,lum[0]);\
 506                     movq_m2r(ptr[16],mm2);\
 507                     movq_m2r(ptr[24],mm1);\
 508                     movq_r2r(mm2,mm4);\
 509                     punpcklbw_r2r(mm1,mm2);\
 510                     punpckhbw_r2r(mm1,mm4);\
 511                     movq_r2r(mm2,mm3);\
 512                     punpcklbw_r2r(mm4,mm2);\
 513                     punpckhbw_r2r(mm4,mm3);\
 514                     movq_r2r(mm2,mm1);\
 515                     movq_r2r(mm3,mm4);\
 516                     punpcklbw_r2r(mm7,mm1);\
 517                     punpcklbw_r2r(mm7,mm4);\
 518                     paddw_r2r(mm6,mm1);\
 519                     paddw_r2r(mm4,mm1);\
 520                     psraw_i2r(1,mm1);\
 521                     packuswb_r2r(mm7,mm1);\
 522                     movd_r2m(mm1,lum[4]);\
 523                     punpckhbw_r2r(mm7,mm0);\
 524                     punpckhbw_r2r(mm7,mm2);\
 525                     paddw_r2r(mm6,mm0);\
 526                     paddw_r2r(mm2,mm0);\
 527                     psraw_i2r(1,mm0);\
 528                     packuswb_r2r(mm7,mm0);\
 529                     punpckhbw_r2r(mm7,mm5);\
 530                     punpckhbw_r2r(mm7,mm3);\
 531                     paddw_r2r(mm6,mm5);\
 532                     paddw_r2r(mm3,mm5);\
 533                     psraw_i2r(1,mm5);\
 534                     packuswb_r2r(mm7,mm5);\
 535                     movd_r2m(mm0,cb[0]);\
 536                     movd_r2m(mm5,cr[0]);
 537
 538 #define LINE_NOUV_AVG \
 539                     movq_m2r(ptr[0],mm0);\
 540                     movq_m2r(ptr[8],mm1);\
 541                     pand_r2r(mm5,mm0);\
 542                     pand_r2r(mm5,mm1);\
 543                     pmaddwd_r2r(mm6,mm0);\
 544                     pmaddwd_r2r(mm6,mm1);\
 545                     packssdw_r2r(mm1,mm0);\
 546                     paddw_r2r(mm6,mm0);\
 547                     psraw_i2r(1,mm0);\
 548                     movq_m2r(ptr[16],mm2);\
 549                     movq_m2r(ptr[24],mm3);\
 550                     pand_r2r(mm5,mm2);\
 551                     pand_r2r(mm5,mm3);\
 552                     pmaddwd_r2r(mm6,mm2);\
 553                     pmaddwd_r2r(mm6,mm3);\
 554                     packssdw_r2r(mm3,mm2);\
 555                     paddw_r2r(mm6,mm2);\
 556                     psraw_i2r(1,mm2);\
 557                     packuswb_r2r(mm2,mm0);\
 558                     movq_r2m(mm0,lum[0]);
 559
 560 #define DEINT_LINE_LUM(ptroff) \
 561                     movd_m2r(lum_m4[(ptroff)],mm0);\
 562                     movd_m2r(lum_m3[(ptroff)],mm1);\
 563                     movd_m2r(lum_m2[(ptroff)],mm2);\
 564                     movd_m2r(lum_m1[(ptroff)],mm3);\
 565                     movd_m2r(lum[(ptroff)],mm4);\
 566                     punpcklbw_r2r(mm7,mm0);\
 567                     movd_r2m(mm2,lum_m4[(ptroff)]);\
 568                     punpcklbw_r2r(mm7,mm1);\
 569                     punpcklbw_r2r(mm7,mm2);\
 570                     punpcklbw_r2r(mm7,mm3);\
 571                     punpcklbw_r2r(mm7,mm4);\
 572                     psllw_i2r(2,mm1);\
 573                     psllw_i2r(1,mm2);\
 574                     paddw_r2r(mm6,mm1);\
 575                     psllw_i2r(2,mm3);\
 576                     paddw_r2r(mm2,mm1);\
 577                     paddw_r2r(mm4,mm0);\
 578                     paddw_r2r(mm3,mm1);\
 579                     psubusw_r2r(mm0,mm1);\
 580                     psrlw_i2r(3,mm1);\
 581                     packuswb_r2r(mm7,mm1);\
 582                     movd_r2m(mm1,lum_m2[(ptroff)]);
 583
 584 #else
 585 #include "dsputil.h"
 586
 587 #define LINE_WITH_UV \
 588                     lum[0]=ptr[0];lum[1]=ptr[2];lum[2]=ptr[4];lum[3]=ptr[6];\
 589                     cb[0]=ptr[1];cb[1]=ptr[5];\
 590                     cr[0]=ptr[3];cr[1]=ptr[7];\
 591                     lum[4]=ptr[8];lum[5]=ptr[10];lum[6]=ptr[12];lum[7]=ptr[14];\
 592                     cb[2]=ptr[9];cb[3]=ptr[13];\
 593                     cr[2]=ptr[11];cr[3]=ptr[15];\
 594                     lum[8]=ptr[16];lum[9]=ptr[18];lum[10]=ptr[20];lum[11]=ptr[22];\
 595                     cb[4]=ptr[17];cb[5]=ptr[21];\
 596                     cr[4]=ptr[19];cr[5]=ptr[23];\
 597                     lum[12]=ptr[24];lum[13]=ptr[26];lum[14]=ptr[28];lum[15]=ptr[30];\
 598                     cb[6]=ptr[25];cb[7]=ptr[29];\
 599                     cr[6]=ptr[27];cr[7]=ptr[31];
 600
 601 #define LINE_NO_UV \
 602                     lum[0]=ptr[0];lum[1]=ptr[2];lum[2]=ptr[4];lum[3]=ptr[6];\
 603                     lum[4]=ptr[8];lum[5]=ptr[10];lum[6]=ptr[12];lum[7]=ptr[14];\
 604                     lum[8]=ptr[16];lum[9]=ptr[18];lum[10]=ptr[20];lum[11]=ptr[22];\
 605                     lum[12]=ptr[24];lum[13]=ptr[26];lum[14]=ptr[28];lum[15]=ptr[30];
 606
 607 #define LINE_WITHUV_AVG \
 608                     sum=(ptr[0]+ptr[2]+1) >> 1;lum[0]=sum; \
 609                     sum=(ptr[4]+ptr[6]+1) >> 1;lum[1]=sum; \
 610                     sum=(ptr[1]+ptr[5]+1) >> 1;cb[0]=sum; \
 611                     sum=(ptr[3]+ptr[7]+1) >> 1;cr[0]=sum; \
 612                     sum=(ptr[8]+ptr[10]+1) >> 1;lum[2]=sum; \
 613                     sum=(ptr[12]+ptr[14]+1) >> 1;lum[3]=sum; \
 614                     sum=(ptr[9]+ptr[13]+1) >> 1;cb[1]=sum; \
 615                     sum=(ptr[11]+ptr[15]+1) >> 1;cr[1]=sum; \
 616                     sum=(ptr[16]+ptr[18]+1) >> 1;lum[4]=sum; \
 617                     sum=(ptr[20]+ptr[22]+1) >> 1;lum[5]=sum; \
 618                     sum=(ptr[17]+ptr[21]+1) >> 1;cb[2]=sum; \
 619                     sum=(ptr[19]+ptr[23]+1) >> 1;cr[2]=sum; \
 620                     sum=(ptr[24]+ptr[26]+1) >> 1;lum[6]=sum; \
 621                     sum=(ptr[28]+ptr[30]+1) >> 1;lum[7]=sum; \
 622                     sum=(ptr[25]+ptr[29]+1) >> 1;cb[3]=sum; \
 623                     sum=(ptr[27]+ptr[31]+1) >> 1;cr[3]=sum;
 624
 625 #define LINE_NOUV_AVG \
 626                     sum=(ptr[0]+ptr[2]+1) >> 1;lum[0]=sum; \
 627                     sum=(ptr[4]+ptr[6]+1) >> 1;lum[1]=sum; \
 628                     sum=(ptr[8]+ptr[10]+1) >> 1;lum[2]=sum; \
 629                     sum=(ptr[12]+ptr[14]+1) >> 1;lum[3]=sum; \
 630                     sum=(ptr[16]+ptr[18]+1) >> 1;lum[4]=sum; \
 631                     sum=(ptr[20]+ptr[22]+1) >> 1;lum[5]=sum; \
 632                     sum=(ptr[24]+ptr[26]+1) >> 1;lum[6]=sum; \
 633                     sum=(ptr[28]+ptr[30]+1) >> 1;lum[7]=sum;
 634
 635 #define DEINT_LINE_LUM(ptroff) \
 636                     sum=(-lum_m4[(ptroff)]+(lum_m3[(ptroff)]<<2)+(lum_m2[(ptroff)]<<1)+(lum_m1[(ptroff)]<<2)-lum[(ptroff)]); \
 637                     lum_m4[(ptroff)]=lum_m2[(ptroff)];\
 638                     lum_m2[(ptroff)]=cm[(sum+4)>>3];\
 639                     sum=(-lum_m4[(ptroff)+1]+(lum_m3[(ptroff)+1]<<2)+(lum_m2[(ptroff)+1]<<1)+(lum_m1[(ptroff)+1]<<2)-lum[(ptroff)+1]); \
 640                     lum_m4[(ptroff)+1]=lum_m2[(ptroff)+1];\
 641                     lum_m2[(ptroff)+1]=cm[(sum+4)>>3];\
 642                     sum=(-lum_m4[(ptroff)+2]+(lum_m3[(ptroff)+2]<<2)+(lum_m2[(ptroff)+2]<<1)+(lum_m1[(ptroff)+2]<<2)-lum[(ptroff)+2]); \
 643                     lum_m4[(ptroff)+2]=lum_m2[(ptroff)+2];\
 644                     lum_m2[(ptroff)+2]=cm[(sum+4)>>3];\
 645                     sum=(-lum_m4[(ptroff)+3]+(lum_m3[(ptroff)+3]<<2)+(lum_m2[(ptroff)+3]<<1)+(lum_m1[(ptroff)+3]<<2)-lum[(ptroff)+3]); \
 646                     lum_m4[(ptroff)+3]=lum_m2[(ptroff)+3];\
 647                     lum_m2[(ptroff)+3]=cm[(sum+4)>>3];
 648
 649 #endif
 650
 651
 652 /* Read two fields separately. */
 653 static int aiw_read_picture(VideoData *s, uint8_t *data)
 654 {
 655     uint8_t *ptr, *lum, *cb, *cr;
 656     int h;
 657 #ifndef HAVE_MMX
 658     int sum;
 659 #endif
 660     uint8_t* src = s->src_mem;
 661     uint8_t *ptrend = &src[s->width*2];
 662     lum=data;
 663     cb=&lum[s->width*s->height];
 664     cr=&cb[(s->width*s->height)/4];
 665     if (s->deint == 0 && s->halfw == 0) {
 666         while (read(s->fd,src,s->width*2) < 0) {
 667             usleep(100);
 668         }
 669         for (h = 0; h < s->height-2; h+=2) {
 670             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
 671                 LINE_WITH_UV
 672                     }
 673             read(s->fd,src,s->width*2);
 674             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16) {
 675                 LINE_NO_UV
 676                     }
 677             read(s->fd,src,s->width*2);
 678         }
 679         /*
 680          * Do last two lines
 681          */
 682         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
 683             LINE_WITH_UV
 684                 }
 685         read(s->fd,src,s->width*2);
 686         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16) {
 687             LINE_NO_UV
 688                 }
 689         /* drop second field */
 690         while (read(s->fd,src,s->width*2) < 0) {
 691             usleep(100);
 692         }
 693         for (h = 0; h < s->height - 1; h++) {
 694             read(s->fd,src,s->width*2);
 695         }
 696     } else if (s->halfw == 1) {
 697 #ifdef HAVE_MMX
 698         mmx_t rounder;
 699         mmx_t masker;
 700         rounder.uw[0]=1;
 701         rounder.uw[1]=1;
 702         rounder.uw[2]=1;
 703         rounder.uw[3]=1;
 704         masker.ub[0]=0xff;
 705         masker.ub[1]=0;
 706         masker.ub[2]=0xff;
 707         masker.ub[3]=0;
 708         masker.ub[4]=0xff;
 709         masker.ub[5]=0;
 710         masker.ub[6]=0xff;
 711         masker.ub[7]=0;
 712         pxor_r2r(mm7,mm7);
 713         movq_m2r(rounder,mm6);
 714 #endif
 715         while (read(s->fd,src,s->width*4) < 0) {
 716             usleep(100);
 717         }
 718         ptrend = &src[s->width*4];
 719         for (h = 0; h < s->height-2; h+=2) {
 720             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=8, cb+=4, cr+=4) {
 721                 LINE_WITHUV_AVG
 722                     }
 723             read(s->fd,src,s->width*4);
 724 #ifdef HAVE_MMX
 725             movq_m2r(masker,mm5);
 726 #endif
 727             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=8) {
 728                 LINE_NOUV_AVG
 729                     }
 730             read(s->fd,src,s->width*4);
 731         }
 732         /*
 733  * Do last two lines
 734  */
 735         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=8, cb+=4, cr+=4) {
 736             LINE_WITHUV_AVG
 737                 }
 738         read(s->fd,src,s->width*4);
 739 #ifdef HAVE_MMX
 740         movq_m2r(masker,mm5);
 741 #endif
 742         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=8) {
 743             LINE_NOUV_AVG
 744                 }
 745         /* drop second field */
 746         while (read(s->fd,src,s->width*4) < 0) {
 747             usleep(100);
 748         }
 749         for (h = 0; h < s->height - 1; h++) {
 750             read(s->fd,src,s->width*4);
 751         }
 752     } else {
 753         uint8_t *lum_m1, *lum_m2, *lum_m3, *lum_m4;
 754 #ifdef HAVE_MMX
 755         mmx_t rounder;
 756         rounder.uw[0]=4;
 757         rounder.uw[1]=4;
 758         rounder.uw[2]=4;
 759         rounder.uw[3]=4;
 760         movq_m2r(rounder,mm6);
 761         pxor_r2r(mm7,mm7);
 762 #else
 763         uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
 764 #endif
 765
 766         /* read two fields and deinterlace them */
 767         while (read(s->fd,src,s->width*2) < 0) {
 768             usleep(100);
 769         }
 770         for (h = 0; h < (s->height/2)-2; h+=2) {
 771             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
 772                 LINE_WITH_UV
 773                     }
 774             read(s->fd,src,s->width*2);
 775             /* skip a luminance line - will be filled in later */
 776             lum += s->width;
 777             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
 778                 LINE_WITH_UV
 779                     }
 780             /* skip a luminance line - will be filled in later */
 781             lum += s->width;
 782             read(s->fd,src,s->width*2);
 783         }
 784         /*
 785  * Do last two lines
 786  */
 787         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
 788             LINE_WITH_UV
 789                 }
 790         /* skip a luminance line - will be filled in later */
 791         lum += s->width;
 792         read(s->fd,src,s->width*2);
 793         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, cb+=8, cr+=8) {
 794             LINE_WITH_UV
 795                 }
 796         /*
 797  *
 798  * SECOND FIELD
 799  *
 800  */
 801         lum=&data[s->width];
 802         while (read(s->fd,src,s->width*2) < 0) {
 803             usleep(10);
 804         }
 805         /* First (and last) two lines not interlaced */
 806         for (h = 0; h < 2; h++) {
 807             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16) {
 808                 LINE_NO_UV
 809                     }
 810             read(s->fd,src,s->width*2);
 811             /* skip a luminance line */
 812             lum += s->width;
 813         }
 814         lum_m1=&lum[-s->width];
 815         lum_m2=&lum_m1[-s->width];
 816         lum_m3=&lum_m2[-s->width];
 817         memmove(s->lum_m4_mem,&lum_m3[-s->width],s->width);
 818         for (; h < (s->height/2)-1; h++) {
 819             lum_m4=s->lum_m4_mem;
 820             for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16,lum_m1+=16,lum_m2+=16,lum_m3+=16,lum_m4+=16) {
 821                 LINE_NO_UV
 822
 823                     DEINT_LINE_LUM(0)
 824                     DEINT_LINE_LUM(4)
 825                     DEINT_LINE_LUM(8)
 826                     DEINT_LINE_LUM(12)
 827                     }
 828             read(s->fd,src,s->width*2);
 829             /* skip a luminance line */
 830             lum += s->width;
 831             lum_m1 += s->width;
 832             lum_m2 += s->width;
 833             lum_m3 += s->width;
 834             //                lum_m4 += s->width;
 835         }
 836         /*
 837  * Do last line
 838  */
 839         lum_m4=s->lum_m4_mem;
 840         for (ptr = &src[0]; ptr < ptrend; ptr+=32, lum+=16, lum_m1+=16, lum_m2+=16, lum_m3+=16, lum_m4+=16) {
 841             LINE_NO_UV
 842
 843                 DEINT_LINE_LUM(0)
 844                 DEINT_LINE_LUM(4)
 845                 DEINT_LINE_LUM(8)
 846                 DEINT_LINE_LUM(12)
 847                 }
 848     }
 849 #ifdef HAVE_MMX
 850     emms();
 851 #endif
 852     return s->frame_size;
 853 }
 854
 855 static int aiw_close(VideoData *s)
 856 {
 857     av_freep(&s->lum_m4_mem);
 858     av_freep(&s->src_mem);
 859     return 0;
 860 }