yading@10: /*
yading@10:  * DV decoder
yading@10:  * Copyright (c) 2002 Fabrice Bellard
yading@10:  * Copyright (c) 2004 Roman Shaposhnik
yading@10:  *
yading@10:  * 50 Mbps (DVCPRO50) support
yading@10:  * Copyright (c) 2006 Daniel Maas <dmaas@maasdigital.com>
yading@10:  *
yading@10:  * 100 Mbps (DVCPRO HD) support
yading@10:  * Initial code by Daniel Maas <dmaas@maasdigital.com> (funded by BBC R&D)
yading@10:  * Final code by Roman Shaposhnik
yading@10:  *
yading@10:  * Many thanks to Dan Dennedy <dan@dennedy.org> for providing wealth
yading@10:  * of DV technical info.
yading@10:  *
yading@10:  * This file is part of FFmpeg.
yading@10:  *
yading@10:  * FFmpeg is free software; you can redistribute it and/or
yading@10:  * modify it under the terms of the GNU Lesser General Public
yading@10:  * License as published by the Free Software Foundation; either
yading@10:  * version 2.1 of the License, or (at your option) any later version.
yading@10:  *
yading@10:  * FFmpeg is distributed in the hope that it will be useful,
yading@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
yading@10:  * Lesser General Public License for more details.
yading@10:  *
yading@10:  * You should have received a copy of the GNU Lesser General Public
yading@10:  * License along with FFmpeg; if not, write to the Free Software
yading@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
yading@10:  */
yading@10: 
yading@10: /**
yading@10:  * @file
yading@10:  * DV decoder
yading@10:  */
yading@10: 
yading@10: #include "libavutil/avassert.h"
yading@10: #include "libavutil/internal.h"
yading@10: #include "libavutil/pixdesc.h"
yading@10: #include "avcodec.h"
yading@10: #include "internal.h"
yading@10: #include "get_bits.h"
yading@10: #include "put_bits.h"
yading@10: #include "simple_idct.h"
yading@10: #include "dvdata.h"
yading@10: 
yading@10: typedef struct BlockInfo {
yading@10:     const uint32_t *factor_table;
yading@10:     const uint8_t *scan_table;
yading@10:     uint8_t pos; /* position in block */
yading@10:     void (*idct_put)(uint8_t *dest, int line_size, int16_t *block);
yading@10:     uint8_t partial_bit_count;
yading@10:     uint32_t partial_bit_buffer;
yading@10:     int shift_offset;
yading@10: } BlockInfo;
yading@10: 
yading@10: static const int dv_iweight_bits = 14;
yading@10: 
yading@10: /* decode AC coefficients */
yading@10: static void dv_decode_ac(GetBitContext *gb, BlockInfo *mb, int16_t *block)
yading@10: {
yading@10:     int last_index = gb->size_in_bits;
yading@10:     const uint8_t  *scan_table   = mb->scan_table;
yading@10:     const uint32_t *factor_table = mb->factor_table;
yading@10:     int pos               = mb->pos;
yading@10:     int partial_bit_count = mb->partial_bit_count;
yading@10:     int level, run, vlc_len, index;
yading@10: 
yading@10:     OPEN_READER(re, gb);
yading@10:     UPDATE_CACHE(re, gb);
yading@10: 
yading@10:     /* if we must parse a partial VLC, we do it here */
yading@10:     if (partial_bit_count > 0) {
yading@10:         re_cache = re_cache >> partial_bit_count | mb->partial_bit_buffer;
yading@10:         re_index -= partial_bit_count;
yading@10:         mb->partial_bit_count = 0;
yading@10:     }
yading@10: 
yading@10:     /* get the AC coefficients until last_index is reached */
yading@10:     for (;;) {
yading@10:         av_dlog(NULL, "%2d: bits=%04x index=%d\n", pos, SHOW_UBITS(re, gb, 16),
yading@10:                 re_index);
yading@10:         /* our own optimized GET_RL_VLC */
yading@10:         index   = NEG_USR32(re_cache, TEX_VLC_BITS);
yading@10:         vlc_len = ff_dv_rl_vlc[index].len;
yading@10:         if (vlc_len < 0) {
yading@10:             index = NEG_USR32((unsigned)re_cache << TEX_VLC_BITS, -vlc_len) +
yading@10:                     ff_dv_rl_vlc[index].level;
yading@10:             vlc_len = TEX_VLC_BITS - vlc_len;
yading@10:         }
yading@10:         level = ff_dv_rl_vlc[index].level;
yading@10:         run   = ff_dv_rl_vlc[index].run;
yading@10: 
yading@10:         /* gotta check if we're still within gb boundaries */
yading@10:         if (re_index + vlc_len > last_index) {
yading@10:             /* should be < 16 bits otherwise a codeword could have been parsed */
yading@10:             mb->partial_bit_count = last_index - re_index;
yading@10:             mb->partial_bit_buffer = re_cache & ~(-1u >> mb->partial_bit_count);
yading@10:             re_index = last_index;
yading@10:             break;
yading@10:         }
yading@10:         re_index += vlc_len;
yading@10: 
yading@10:         av_dlog(NULL, "run=%d level=%d\n", run, level);
yading@10:         pos += run;
yading@10:         if (pos >= 64)
yading@10:             break;
yading@10: 
yading@10:         level = (level * factor_table[pos] + (1 << (dv_iweight_bits - 1))) >> dv_iweight_bits;
yading@10:         block[scan_table[pos]] = level;
yading@10: 
yading@10:         UPDATE_CACHE(re, gb);
yading@10:     }
yading@10:     CLOSE_READER(re, gb);
yading@10:     mb->pos = pos;
yading@10: }
yading@10: 
yading@10: static inline void bit_copy(PutBitContext *pb, GetBitContext *gb)
yading@10: {
yading@10:     int bits_left = get_bits_left(gb);
yading@10:     while (bits_left >= MIN_CACHE_BITS) {
yading@10:         put_bits(pb, MIN_CACHE_BITS, get_bits(gb, MIN_CACHE_BITS));
yading@10:         bits_left -= MIN_CACHE_BITS;
yading@10:     }
yading@10:     if (bits_left > 0) {
yading@10:         put_bits(pb, bits_left, get_bits(gb, bits_left));
yading@10:     }
yading@10: }
yading@10: 
yading@10: /* mb_x and mb_y are in units of 8 pixels */
yading@10: static int dv_decode_video_segment(AVCodecContext *avctx, void *arg)
yading@10: {
yading@10:     DVVideoContext *s = avctx->priv_data;
yading@10:     DVwork_chunk *work_chunk = arg;
yading@10:     int quant, dc, dct_mode, class1, j;
yading@10:     int mb_index, mb_x, mb_y, last_index;
yading@10:     int y_stride, linesize;
yading@10:     int16_t *block, *block1;
yading@10:     int c_offset;
yading@10:     uint8_t *y_ptr;
yading@10:     const uint8_t *buf_ptr;
yading@10:     PutBitContext pb, vs_pb;
yading@10:     GetBitContext gb;
yading@10:     BlockInfo mb_data[5 * DV_MAX_BPM], *mb, *mb1;
yading@10:     LOCAL_ALIGNED_16(int16_t, sblock, [5*DV_MAX_BPM], [64]);
yading@10:     LOCAL_ALIGNED_16(uint8_t, mb_bit_buffer, [  80 + FF_INPUT_BUFFER_PADDING_SIZE]); /* allow some slack */
yading@10:     LOCAL_ALIGNED_16(uint8_t, vs_bit_buffer, [5*80 + FF_INPUT_BUFFER_PADDING_SIZE]); /* allow some slack */
yading@10:     const int log2_blocksize = 3-s->avctx->lowres;
yading@10:     int is_field_mode[5];
yading@10: 
yading@10:     av_assert1((((int)mb_bit_buffer) & 7) == 0);
yading@10:     av_assert1((((int)vs_bit_buffer) & 7) == 0);
yading@10: 
yading@10:     memset(sblock, 0, 5*DV_MAX_BPM*sizeof(*sblock));
yading@10: 
yading@10:     /* pass 1: read DC and AC coefficients in blocks */
yading@10:     buf_ptr = &s->buf[work_chunk->buf_offset*80];
yading@10:     block1  = &sblock[0][0];
yading@10:     mb1     = mb_data;
yading@10:     init_put_bits(&vs_pb, vs_bit_buffer, 5 * 80);
yading@10:     for (mb_index = 0; mb_index < 5; mb_index++, mb1 += s->sys->bpm, block1 += s->sys->bpm * 64) {
yading@10:         /* skip header */
yading@10:         quant = buf_ptr[3] & 0x0f;
yading@10:         buf_ptr += 4;
yading@10:         init_put_bits(&pb, mb_bit_buffer, 80);
yading@10:         mb    = mb1;
yading@10:         block = block1;
yading@10:         is_field_mode[mb_index] = 0;
yading@10:         for (j = 0; j < s->sys->bpm; j++) {
yading@10:             last_index = s->sys->block_sizes[j];
yading@10:             init_get_bits(&gb, buf_ptr, last_index);
yading@10: 
yading@10:             /* get the DC */
yading@10:             dc       = get_sbits(&gb, 9);
yading@10:             dct_mode = get_bits1(&gb);
yading@10:             class1   = get_bits(&gb, 2);
yading@10:             if (DV_PROFILE_IS_HD(s->sys)) {
yading@10:                 mb->idct_put     = s->idct_put[0];
yading@10:                 mb->scan_table   = s->dv_zigzag[0];
yading@10:                 mb->factor_table = &s->sys->idct_factor[(j >= 4)*4*16*64 + class1*16*64 + quant*64];
yading@10:                 is_field_mode[mb_index] |= !j && dct_mode;
yading@10:             } else {
yading@10:                 mb->idct_put     = s->idct_put[dct_mode && log2_blocksize == 3];
yading@10:                 mb->scan_table   = s->dv_zigzag[dct_mode];
yading@10:                 mb->factor_table = &s->sys->idct_factor[(class1 == 3)*2*22*64 + dct_mode*22*64 +
yading@10:                                                         (quant + ff_dv_quant_offset[class1])*64];
yading@10:             }
yading@10:             dc = dc << 2;
yading@10:             /* convert to unsigned because 128 is not added in the
yading@10:                standard IDCT */
yading@10:             dc += 1024;
yading@10:             block[0] = dc;
yading@10:             buf_ptr += last_index >> 3;
yading@10:             mb->pos               = 0;
yading@10:             mb->partial_bit_count = 0;
yading@10: 
yading@10:             av_dlog(avctx, "MB block: %d, %d ", mb_index, j);
yading@10:             dv_decode_ac(&gb, mb, block);
yading@10: 
yading@10:             /* write the remaining bits in a new buffer only if the
yading@10:                block is finished */
yading@10:             if (mb->pos >= 64)
yading@10:                 bit_copy(&pb, &gb);
yading@10: 
yading@10:             block += 64;
yading@10:             mb++;
yading@10:         }
yading@10: 
yading@10:         /* pass 2: we can do it just after */
yading@10:         av_dlog(avctx, "***pass 2 size=%d MB#=%d\n", put_bits_count(&pb), mb_index);
yading@10:         block = block1;
yading@10:         mb    = mb1;
yading@10:         init_get_bits(&gb, mb_bit_buffer, put_bits_count(&pb));
yading@10:         put_bits32(&pb, 0); // padding must be zeroed
yading@10:         flush_put_bits(&pb);
yading@10:         for (j = 0; j < s->sys->bpm; j++, block += 64, mb++) {
yading@10:             if (mb->pos < 64 && get_bits_left(&gb) > 0) {
yading@10:                 dv_decode_ac(&gb, mb, block);
yading@10:                 /* if still not finished, no need to parse other blocks */
yading@10:                 if (mb->pos < 64)
yading@10:                     break;
yading@10:             }
yading@10:         }
yading@10:         /* all blocks are finished, so the extra bytes can be used at
yading@10:            the video segment level */
yading@10:         if (j >= s->sys->bpm)
yading@10:             bit_copy(&vs_pb, &gb);
yading@10:     }
yading@10: 
yading@10:     /* we need a pass over the whole video segment */
yading@10:     av_dlog(avctx, "***pass 3 size=%d\n", put_bits_count(&vs_pb));
yading@10:     block = &sblock[0][0];
yading@10:     mb    = mb_data;
yading@10:     init_get_bits(&gb, vs_bit_buffer, put_bits_count(&vs_pb));
yading@10:     put_bits32(&vs_pb, 0); // padding must be zeroed
yading@10:     flush_put_bits(&vs_pb);
yading@10:     for (mb_index = 0; mb_index < 5; mb_index++) {
yading@10:         for (j = 0; j < s->sys->bpm; j++) {
yading@10:             if (mb->pos < 64) {
yading@10:                 av_dlog(avctx, "start %d:%d\n", mb_index, j);
yading@10:                 dv_decode_ac(&gb, mb, block);
yading@10:             }
yading@10:             if (mb->pos >= 64 && mb->pos < 127)
yading@10:                 av_log(avctx, AV_LOG_ERROR, "AC EOB marker is absent pos=%d\n", mb->pos);
yading@10:             block += 64;
yading@10:             mb++;
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     /* compute idct and place blocks */
yading@10:     block = &sblock[0][0];
yading@10:     mb    = mb_data;
yading@10:     for (mb_index = 0; mb_index < 5; mb_index++) {
yading@10:         dv_calculate_mb_xy(s, work_chunk, mb_index, &mb_x, &mb_y);
yading@10: 
yading@10:         /* idct_put'ting luminance */
yading@10:         if ((s->sys->pix_fmt == AV_PIX_FMT_YUV420P) ||
yading@10:             (s->sys->pix_fmt == AV_PIX_FMT_YUV411P && mb_x >= (704 / 8)) ||
yading@10:             (s->sys->height >= 720 && mb_y != 134)) {
yading@10:             y_stride = (s->picture.linesize[0] << ((!is_field_mode[mb_index]) * log2_blocksize));
yading@10:         } else {
yading@10:             y_stride = (2 << log2_blocksize);
yading@10:         }
yading@10:         y_ptr = s->picture.data[0] + ((mb_y * s->picture.linesize[0] + mb_x) << log2_blocksize);
yading@10:         linesize = s->picture.linesize[0] << is_field_mode[mb_index];
yading@10:         mb[0]    .idct_put(y_ptr                                   , linesize, block + 0*64);
yading@10:         if (s->sys->video_stype == 4) { /* SD 422 */
yading@10:             mb[2].idct_put(y_ptr + (1 << log2_blocksize)           , linesize, block + 2*64);
yading@10:         } else {
yading@10:             mb[1].idct_put(y_ptr + (1 << log2_blocksize)           , linesize, block + 1*64);
yading@10:             mb[2].idct_put(y_ptr                         + y_stride, linesize, block + 2*64);
yading@10:             mb[3].idct_put(y_ptr + (1 << log2_blocksize) + y_stride, linesize, block + 3*64);
yading@10:         }
yading@10:         mb += 4;
yading@10:         block += 4*64;
yading@10: 
yading@10:         /* idct_put'ting chrominance */
yading@10:         c_offset = (((mb_y >>  (s->sys->pix_fmt == AV_PIX_FMT_YUV420P)) * s->picture.linesize[1] +
yading@10:                      (mb_x >> ((s->sys->pix_fmt == AV_PIX_FMT_YUV411P) ? 2 : 1))) << log2_blocksize);
yading@10:         for (j = 2; j; j--) {
yading@10:             uint8_t *c_ptr = s->picture.data[j] + c_offset;
yading@10:             if (s->sys->pix_fmt == AV_PIX_FMT_YUV411P && mb_x >= (704 / 8)) {
yading@10:                   uint64_t aligned_pixels[64/8];
yading@10:                   uint8_t *pixels = (uint8_t*)aligned_pixels;
yading@10:                   uint8_t *c_ptr1, *ptr1;
yading@10:                   int x, y;
yading@10:                   mb->idct_put(pixels, 8, block);
yading@10:                   for (y = 0; y < (1 << log2_blocksize); y++, c_ptr += s->picture.linesize[j], pixels += 8) {
yading@10:                       ptr1   = pixels + ((1 << (log2_blocksize))>>1);
yading@10:                       c_ptr1 = c_ptr + (s->picture.linesize[j] << log2_blocksize);
yading@10:                       for (x = 0; x < (1 << FFMAX(log2_blocksize - 1, 0)); x++) {
yading@10:                           c_ptr[x]  = pixels[x];
yading@10:                           c_ptr1[x] = ptr1[x];
yading@10:                       }
yading@10:                   }
yading@10:                   block += 64; mb++;
yading@10:             } else {
yading@10:                   y_stride = (mb_y == 134) ? (1 << log2_blocksize) :
yading@10:                                              s->picture.linesize[j] << ((!is_field_mode[mb_index]) * log2_blocksize);
yading@10:                   linesize = s->picture.linesize[j] << is_field_mode[mb_index];
yading@10:                   (mb++)->    idct_put(c_ptr           , linesize, block); block += 64;
yading@10:                   if (s->sys->bpm == 8) {
yading@10:                       (mb++)->idct_put(c_ptr + y_stride, linesize, block); block += 64;
yading@10:                   }
yading@10:             }
yading@10:         }
yading@10:     }
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: /* NOTE: exactly one frame must be given (120000 bytes for NTSC,
yading@10:    144000 bytes for PAL - or twice those for 50Mbps) */
yading@10: static int dvvideo_decode_frame(AVCodecContext *avctx,
yading@10:                                  void *data, int *got_frame,
yading@10:                                  AVPacket *avpkt)
yading@10: {
yading@10:     uint8_t *buf = avpkt->data;
yading@10:     int buf_size = avpkt->size;
yading@10:     DVVideoContext *s = avctx->priv_data;
yading@10:     const uint8_t* vsc_pack;
yading@10:     int ret, apt, is16_9;
yading@10: 
yading@10:     s->sys = avpriv_dv_frame_profile2(avctx, s->sys, buf, buf_size);
yading@10:     if (!s->sys || buf_size < s->sys->frame_size || ff_dv_init_dynamic_tables(s->sys)) {
yading@10:         av_log(avctx, AV_LOG_ERROR, "could not find dv frame profile\n");
yading@10:         return -1; /* NOTE: we only accept several full frames */
yading@10:     }
yading@10: 
yading@10:     s->picture.key_frame = 1;
yading@10:     s->picture.pict_type = AV_PICTURE_TYPE_I;
yading@10:     avctx->pix_fmt   = s->sys->pix_fmt;
yading@10:     avctx->time_base = s->sys->time_base;
yading@10:     avcodec_set_dimensions(avctx, s->sys->width, s->sys->height);
yading@10:     if ((ret = ff_get_buffer(avctx, &s->picture, 0)) < 0)
yading@10:         return ret;
yading@10:     s->picture.interlaced_frame = 1;
yading@10:     s->picture.top_field_first  = 0;
yading@10: 
yading@10:     /* Determine the codec's sample_aspect ratio and field order from the packet */
yading@10:     vsc_pack = buf + 80*5 + 48 + 5;
yading@10:     if ( *vsc_pack == dv_video_control ) {
yading@10:         apt = buf[4] & 0x07;
yading@10:         is16_9 = (vsc_pack[2] & 0x07) == 0x02 || (!apt && (vsc_pack[2] & 0x07) == 0x07);
yading@10:         avctx->sample_aspect_ratio = s->sys->sar[is16_9];
yading@10:         s->picture.top_field_first = !(vsc_pack[3] & 0x40);
yading@10:     }
yading@10: 
yading@10:     s->buf = buf;
yading@10:     avctx->execute(avctx, dv_decode_video_segment, s->sys->work_chunks, NULL,
yading@10:                    dv_work_pool_size(s->sys), sizeof(DVwork_chunk));
yading@10: 
yading@10:     emms_c();
yading@10: 
yading@10:     /* return image */
yading@10:     *got_frame = 1;
yading@10:     av_frame_move_ref(data, &s->picture);
yading@10: 
yading@10:     return s->sys->frame_size;
yading@10: }
yading@10: 
yading@10: static int dvvideo_close(AVCodecContext *c)
yading@10: {
yading@10:     DVVideoContext *s = c->priv_data;
yading@10: 
yading@10:     av_frame_unref(&s->picture);
yading@10: 
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: AVCodec ff_dvvideo_decoder = {
yading@10:     .name           = "dvvideo",
yading@10:     .type           = AVMEDIA_TYPE_VIDEO,
yading@10:     .id             = AV_CODEC_ID_DVVIDEO,
yading@10:     .priv_data_size = sizeof(DVVideoContext),
yading@10:     .init           = ff_dvvideo_init,
yading@10:     .close          = dvvideo_close,
yading@10:     .decode         = dvvideo_decode_frame,
yading@10:     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_SLICE_THREADS,
yading@10:     .max_lowres     = 3,
yading@10:     .long_name      = NULL_IF_CONFIG_SMALL("DV (Digital Video)"),
yading@10: };