[FFmpeg-devel] [PATCH] avformat/flacdec: Implement decoding of 32-bit PCM

Martijn van Beurden mvanb1 at gmail.com
Mon Jul 11 22:02:54 EEST 2022


Op ma 20 jun. 2022 om 22:14 schreef Martijn van Beurden <mvanb1 at gmail.com>:

> Recently libFLAC gained the ability (not in any released version yet
> though) to create FLAC files containing 32-bit int PCM samples. To
> keep complexity reasonable, the choice was made to limit residuals
> to 32-bit integers, which the encoder must make sure of. In case
> the encoder cannot find any predictor of which the residuals fit
> this limit, it must default to using a verbatim subframe. Tests have
> shown that this does not happen often (<0.1% of subframes on a
> music corpus of various styles). See also discussion here:
> https://github.com/ietf-wg-cellar/flac-specification/pull/148
>
> This patch adds decoding of these files to ffmpeg.
> ---
>  libavcodec/flac.c     |   4 +-
>  libavcodec/flacdec.c  | 248 ++++++++++++++++++++++++++++++++++++++----
>  libavcodec/get_bits.h |  12 ++
>  libavcodec/mathops.h  |   9 ++
>  4 files changed, 250 insertions(+), 23 deletions(-)
>
> diff --git a/libavcodec/flac.c b/libavcodec/flac.c
> index dd68830622..f326d8fa5c 100644
> --- a/libavcodec/flac.c
> +++ b/libavcodec/flac.c
> @@ -27,7 +27,7 @@
>  #include "flac.h"
>  #include "flacdata.h"
>
> -static const int8_t sample_size_table[] = { 0, 8, 12, 0, 16, 20, 24, 0 };
> +static const int8_t sample_size_table[] = { 0, 8, 12, 0, 16, 20, 24, 32 };
>
>  static const AVChannelLayout flac_channel_layouts[8] = {
>      AV_CHANNEL_LAYOUT_MONO,
> @@ -81,7 +81,7 @@ int ff_flac_decode_frame_header(AVCodecContext *avctx,
> GetBitContext *gb,
>
>      /* bits per sample */
>      bps_code = get_bits(gb, 3);
> -    if (bps_code == 3 || bps_code == 7) {
> +    if (bps_code == 3) {
>          av_log(avctx, AV_LOG_ERROR + log_level_offset,
>                 "invalid sample size code (%d)\n",
>                 bps_code);
> diff --git a/libavcodec/flacdec.c b/libavcodec/flacdec.c
> index 87f20c7425..49952ce120 100644
> --- a/libavcodec/flacdec.c
> +++ b/libavcodec/flacdec.c
> @@ -63,6 +63,9 @@ typedef struct FLACContext {
>      int32_t *decoded[FLAC_MAX_CHANNELS];    ///< decoded samples
>      uint8_t *decoded_buffer;
>      unsigned int decoded_buffer_size;
> +    int64_t *decoded_33bps;                  ///< decoded samples for a
> 33 bps subframe
> +    uint8_t *decoded_buffer_33bps;
> +    unsigned int decoded_buffer_size_33bps;
>      int buggy_lpc;                          ///< use workaround for old
> lavc encoded files
>
>      FLACDSPContext dsp;
> @@ -154,6 +157,24 @@ static int allocate_buffers(FLACContext *s)
>                                   s->flac_stream_info.channels,
>                                   s->flac_stream_info.max_blocksize,
>                                   AV_SAMPLE_FMT_S32P, 0);
> +    if (ret >= 0 && s->flac_stream_info.bps == 32 &&
> s->flac_stream_info.channels == 2) {
> +        buf_size = av_samples_get_buffer_size(NULL, 1,
> +
> s->flac_stream_info.max_blocksize,
> +                                              AV_SAMPLE_FMT_S64P, 0);
> +        if (buf_size < 0)
> +            return buf_size;
> +
> +        av_fast_malloc(&s->decoded_buffer_33bps,
> &s->decoded_buffer_size_33bps, buf_size);
> +        if (!s->decoded_buffer)
> +            return AVERROR(ENOMEM);
> +
> +        ret = av_samples_fill_arrays((uint8_t **)&s->decoded_33bps, NULL,
> +                                     s->decoded_buffer_33bps,
> +                                     1,
> +                                     s->flac_stream_info.max_blocksize,
> +                                     AV_SAMPLE_FMT_S64P, 0);
> +
> +    }
>      return ret < 0 ? ret : 0;
>  }
>
> @@ -331,6 +352,94 @@ static int decode_subframe_fixed(FLACContext *s,
> int32_t *decoded,
>      return 0;
>  }
>
> +static int decode_subframe_fixed_wide(FLACContext *s, int32_t *decoded,
> +                                      int pred_order, int bps)
> +{
> +    const int blocksize = s->blocksize;
> +    int i;
> +    int ret;
> +
> +    /* warm up samples */
> +    for (i = 0; i < pred_order; i++) {
> +        decoded[i] = get_sbits_long(&s->gb, bps);
> +    }
> +
> +    if ((ret = decode_residuals(s, decoded, pred_order)) < 0)
> +        return ret;
> +
> +    switch (pred_order) {
> +    case 0:
> +        break;
> +    case 1:
> +        for (i = pred_order; i < blocksize; i++)
> +            decoded[i] += decoded[i-1];
> +        break;
> +    case 2:
> +        for (i = pred_order; i < blocksize; i++)
> +            decoded[i] = (int64_t)decoded[i] + 2*(int64_t)decoded[i-1] -
> (int64_t)decoded[i-2];
> +        break;
> +    case 3:
> +        for (i = pred_order; i < blocksize; i++)
> +            decoded[i] = (int64_t)decoded[i] + 3*(int64_t)decoded[i-1] -
> 3*(int64_t)decoded[i-2] + (int64_t)decoded[i-3];
> +        break;
> +    case 4:
> +        for (i = pred_order; i < blocksize; i++)
> +            decoded[i] = (int64_t)decoded[i] + 4*(int64_t)decoded[i-1] -
> 6*(int64_t)decoded[i-2] + 4*(int64_t)decoded[i-3] - (int64_t)decoded[i-4];
> +        break;
> +    default:
> +        av_log(s->avctx, AV_LOG_ERROR, "illegal pred order %d\n",
> pred_order);
> +        return AVERROR_INVALIDDATA;
> +    }
> +
> +    return 0;
> +}
> +
> +
> +static int decode_subframe_fixed_33bps(FLACContext *s, int64_t *decoded,
> +                                       int32_t *residual, int pred_order)
> +{
> +    const int blocksize = s->blocksize;
> +    int i;
> +    int ret;
> +
> +    /* warm up samples */
> +    for (i = 0; i < pred_order; i++) {
> +        decoded[i] = get_sbits64(&s->gb, 33);
> +        av_log(s->avctx, AV_LOG_DEBUG, "warm-up %d = %" PRId64 "\n", i,
> decoded[i]);
> +    }
> +
> +    if ((ret = decode_residuals(s, residual, pred_order)) < 0)
> +        return ret;
> +
> +    switch (pred_order) {
> +    case 0:
> +        for (i = pred_order; i < blocksize; i++)
> +            decoded[i] = residual[i];
> +        break;
> +    case 1:
> +        for (i = pred_order; i < blocksize; i++)
> +            decoded[i] = residual[i] + decoded[i-1];
> +        break;
> +    case 2:
> +        for (i = pred_order; i < blocksize; i++)
> +            decoded[i] = residual[i] + 2*decoded[i-1] - decoded[i-2];
> +        break;
> +    case 3:
> +        for (i = pred_order; i < blocksize; i++)
> +            decoded[i] = residual[i] + 3*decoded[i-1] - 3*decoded[i-2] +
> decoded[i-3];
> +        break;
> +    case 4:
> +        for (i = pred_order; i < blocksize; i++)
> +            decoded[i] = residual[i] + 4*decoded[i-1] - 6*decoded[i-2] +
> 4*decoded[i-3] - decoded[i-4];
> +        break;
> +    default:
> +        av_log(s->avctx, AV_LOG_ERROR, "illegal pred order %d\n",
> pred_order);
> +        return AVERROR_INVALIDDATA;
> +    }
> +
> +    return 0;
> +}
> +
>  static void lpc_analyze_remodulate(SUINT32 *decoded, const int coeffs[32],
>                                     int order, int qlevel, int len, int
> bps)
>  {
> @@ -402,12 +511,53 @@ static int decode_subframe_lpc(FLACContext *s,
> int32_t *decoded, int pred_order,
>      return 0;
>  }
>
> +static int decode_subframe_lpc_33bps(FLACContext *s, int64_t *decoded,
> +                                     int32_t *residual, int pred_order)
> +{
> +    int i, j, ret;
> +    int coeff_prec, qlevel;
> +    int coeffs[32];
> +
> +    /* warm up samples */
> +    for (i = 0; i < pred_order; i++) {
> +        decoded[i] = get_sbits64(&s->gb, 33);
> +    }
> +
> +    coeff_prec = get_bits(&s->gb, 4) + 1;
> +    if (coeff_prec == 16) {
> +        av_log(s->avctx, AV_LOG_ERROR, "invalid coeff precision\n");
> +        return AVERROR_INVALIDDATA;
> +    }
> +    qlevel = get_sbits(&s->gb, 5);
> +    if (qlevel < 0) {
> +        av_log(s->avctx, AV_LOG_ERROR, "qlevel %d not supported, maybe
> buggy stream\n",
> +               qlevel);
> +        return AVERROR_INVALIDDATA;
> +    }
> +
> +    for (i = 0; i < pred_order; i++) {
> +        coeffs[pred_order - i - 1] = get_sbits(&s->gb, coeff_prec);
> +    }
> +
> +    if ((ret = decode_residuals(s, residual, pred_order)) < 0)
> +        return ret;
> +
> +    for (i = pred_order; i < s->blocksize; i++, decoded++) {
> +        int64_t sum = 0;
> +        for (j = 0; j < pred_order; j++)
> +            sum += (int64_t)coeffs[j] * decoded[j];
> +        decoded[j] = residual[i] + (sum >> qlevel);
> +    }
> +
> +    return 0;
> +}
> +
>  static inline int decode_subframe(FLACContext *s, int channel)
>  {
>      int32_t *decoded = s->decoded[channel];
>      int type, wasted = 0;
>      int bps = s->flac_stream_info.bps;
> -    int i, tmp, ret;
> +    int i, ret;
>
>      if (channel == 0) {
>          if (s->ch_mode == FLAC_CHMODE_RIGHT_SIDE)
> @@ -436,34 +586,63 @@ static inline int decode_subframe(FLACContext *s,
> int channel)
>          wasted = 1 + get_unary(&s->gb, 1, get_bits_left(&s->gb));
>          bps -= wasted;
>      }
> -    if (bps > 32) {
> -        avpriv_report_missing_feature(s->avctx, "Decorrelated bit depth >
> 32");
> -        return AVERROR_PATCHWELCOME;
> -    }
>
>  //FIXME use av_log2 for types
>      if (type == 0) {
> -        tmp = get_sbits_long(&s->gb, bps);
> -        for (i = 0; i < s->blocksize; i++)
> -            decoded[i] = tmp;
> +        if (bps < 33) {
> +            int32_t tmp = get_sbits_long(&s->gb, bps);
> +            for (i = 0; i < s->blocksize; i++)
> +                decoded[i] = tmp;
> +        } else {
> +            int64_t tmp = get_sbits64(&s->gb, 33);
> +            for (i = 0; i < s->blocksize; i++)
> +                s->decoded_33bps[i] = tmp;
> +        }
>      } else if (type == 1) {
> -        for (i = 0; i < s->blocksize; i++)
> -            decoded[i] = get_sbits_long(&s->gb, bps);
> +        if (bps < 33) {
> +            for (i = 0; i < s->blocksize; i++)
> +                decoded[i] = get_sbits_long(&s->gb, bps);
> +        } else {
> +            for (i = 0; i < s->blocksize; i++)
> +                s->decoded_33bps[i] = get_sbits64(&s->gb, 33);
> +        }
>      } else if ((type >= 8) && (type <= 12)) {
> -        if ((ret = decode_subframe_fixed(s, decoded, type & ~0x8, bps)) <
> 0)
> -            return ret;
> +        int order = type & ~0x8;
> +        if (bps < 33) {
> +            if (bps + order <= 32) {
> +                if ((ret = decode_subframe_fixed(s, decoded, order, bps))
> < 0)
> +                    return ret;
> +            } else {
> +                if ((ret = decode_subframe_fixed_wide(s, decoded, order,
> bps)) < 0)
> +                    return ret;
> +            }
> +        } else {
> +            if ((ret = decode_subframe_fixed_33bps(s, s->decoded_33bps,
> decoded, order)) < 0)
> +                return ret;
> +        }
>      } else if (type >= 32) {
> -        if ((ret = decode_subframe_lpc(s, decoded, (type & ~0x20)+1,
> bps)) < 0)
> -            return ret;
> +        if (bps < 33) {
> +            if ((ret = decode_subframe_lpc(s, decoded, (type & ~0x20)+1,
> bps)) < 0)
> +                return ret;
> +        } else {
> +            if ((ret = decode_subframe_lpc_33bps(s, s->decoded_33bps,
> decoded, (type & ~0x20)+1)) < 0)
> +                return ret;
> +        }
>      } else {
>          av_log(s->avctx, AV_LOG_ERROR, "invalid coding type\n");
>          return AVERROR_INVALIDDATA;
>      }
>
> -    if (wasted && wasted < 32) {
> -        int i;
> -        for (i = 0; i < s->blocksize; i++)
> -            decoded[i] = (unsigned)decoded[i] << wasted;
> +    if (wasted) {
> +        if (wasted+bps == 33) {
> +            int i;
> +            for (i = 0; i < s->blocksize; i++)
> +                s->decoded_33bps[i] = (uint64_t)decoded[i] << wasted;
> +        } else if (wasted < 32) {
> +            int i;
> +            for (i = 0; i < s->blocksize; i++)
> +                decoded[i] = (unsigned)decoded[i] << wasted;
> +        }
>      }
>
>      return 0;
> @@ -554,6 +733,26 @@ static int decode_frame(FLACContext *s)
>      return 0;
>  }
>
> +static void decorrelate_33bps(int ch_mode, int32_t **decoded, int64_t
> *decoded_33bps, int len)
> +{
> +    int i;
> +    if (ch_mode == FLAC_CHMODE_LEFT_SIDE ) {
> +        for (i = 0; i < len; i++)
> +           decoded[1][i] = decoded[0][i] - decoded_33bps[i];
> +    } else if (ch_mode == FLAC_CHMODE_RIGHT_SIDE ) {
> +        for (i = 0; i < len; i++)
> +           decoded[0][i] = decoded[1][i] + decoded_33bps[i];
> +    } else if (ch_mode == FLAC_CHMODE_MID_SIDE ) {
> +        for (i = 0; i < len; i++) {
> +            uint64_t a = decoded[0][i];
> +            int64_t b = decoded_33bps[i];
> +            a -= b >> 1;
> +            decoded[0][i] = (a + b);
> +            decoded[1][i] = a;
> +        }
> +    }
> +}
> +
>  static int flac_decode_frame(AVCodecContext *avctx, AVFrame *frame,
>                               int *got_frame_ptr, AVPacket *avpkt)
>  {
> @@ -618,9 +817,15 @@ static int flac_decode_frame(AVCodecContext *avctx,
> AVFrame *frame,
>      if ((ret = ff_thread_get_buffer(avctx, frame, 0)) < 0)
>          return ret;
>
> -    s->dsp.decorrelate[s->ch_mode](frame->data, s->decoded,
> -                                   s->flac_stream_info.channels,
> -                                   s->blocksize, s->sample_shift);
> +    if (s->flac_stream_info.bps == 32 && s->ch_mode > 0) {
> +        decorrelate_33bps(s->ch_mode, s->decoded, s->decoded_33bps,
> s->blocksize);
> +        s->dsp.decorrelate[0](frame->data, s->decoded,
> s->flac_stream_info.channels,
> +                              s->blocksize, s->sample_shift);
> +    } else {
> +        s->dsp.decorrelate[s->ch_mode](frame->data, s->decoded,
> +                                       s->flac_stream_info.channels,
> +                                       s->blocksize, s->sample_shift);
> +    }
>
>      if (bytes_read > buf_size) {
>          av_log(s->avctx, AV_LOG_ERROR, "overread: %d\n", bytes_read -
> buf_size);
> @@ -641,6 +846,7 @@ static av_cold int flac_decode_close(AVCodecContext
> *avctx)
>      FLACContext *s = avctx->priv_data;
>
>      av_freep(&s->decoded_buffer);
> +    av_freep(&s->decoded_buffer_33bps);
>
>      return 0;
>  }
> diff --git a/libavcodec/get_bits.h b/libavcodec/get_bits.h
> index 16f8af5107..c12b62d4be 100644
> --- a/libavcodec/get_bits.h
> +++ b/libavcodec/get_bits.h
> @@ -597,6 +597,18 @@ static inline int get_sbits_long(GetBitContext *s,
> int n)
>      return sign_extend(get_bits_long(s, n), n);
>  }
>
> +/**
> + * Read 0-64 bits as a signed integer.
> + */
> +static inline int64_t get_sbits64(GetBitContext *s, int n)
> +{
> +    // sign_extend(x, 0) is undefined
> +    if (!n)
> +        return 0;
> +
> +    return sign_extend64(get_bits64(s, n), n);
> +}
> +
>  /**
>   * Show 0-32 bits.
>   */
> diff --git a/libavcodec/mathops.h b/libavcodec/mathops.h
> index f81d21f9c4..8a82d9d086 100644
> --- a/libavcodec/mathops.h
> +++ b/libavcodec/mathops.h
> @@ -137,6 +137,15 @@ static inline av_const int sign_extend(int val,
> unsigned bits)
>  }
>  #endif
>
> +#ifndef sign_extend64
> +static inline av_const int64_t sign_extend64(int64_t val, unsigned bits)
> +{
> +    unsigned shift = 8 * sizeof(int64_t) - bits;
> +    union { uint64_t u; int64_t s; } v = { (uint64_t) val << shift };
> +    return v.s >> shift;
> +}
> +#endif
> +
>  #ifndef zero_extend
>  static inline av_const unsigned zero_extend(unsigned val, unsigned bits)
>  {
> --
> 2.30.2
>
>
With this I'd like to once more bring this patch to the attention of the
mailinglist.

Additionally. here's a file to test with:
http://www.audiograaf.nl/misc_stuff/32-bit-test-FLAC.flac


More information about the ffmpeg-devel mailing list