[FFmpeg-devel] [PATCH v8 1/6] lavu/frame: Add Dolby Vision metadata side data type

Andreas Rheinhardt andreas.rheinhardt at outlook.com
Wed Dec 22 16:29:35 EET 2021


Hendrik Leppkes:
> On Wed, Dec 22, 2021 at 2:33 PM Andreas Rheinhardt
> <andreas.rheinhardt at outlook.com> wrote:
>>
>> Hendrik Leppkes:
>>> On Wed, Dec 22, 2021 at 2:00 PM Lynne <dev at lynne.ee> wrote:
>>>>
>>>> 20 Dec 2021, 16:31 by ffmpeg at haasn.xyz:
>>>>
>>>>> From: Niklas Haas <git at haasn.dev>
>>>>>
>>>>> Signed-off-by: Niklas Haas <git at haasn.dev>
>>>>> ---
>>>>>  doc/APIchanges        |   3 +
>>>>>  libavutil/dovi_meta.c |  12 ++++
>>>>>  libavutil/dovi_meta.h | 143 ++++++++++++++++++++++++++++++++++++++++++
>>>>>  libavutil/frame.c     |   1 +
>>>>>  libavutil/frame.h     |   9 ++-
>>>>>  libavutil/version.h   |   2 +-
>>>>>  6 files changed, 168 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/doc/APIchanges b/doc/APIchanges
>>>>> index 17aa664ca3..ff78edec88 100644
>>>>> --- a/doc/APIchanges
>>>>> +++ b/doc/APIchanges
>>>>> @@ -14,6 +14,9 @@ libavutil:     2021-04-27
>>>>>
>>>>>  API changes, most recent first:
>>>>>
>>>>> +2021-12-xx - xxxxxxxxxx - lavu 57.12.100 - frame.h
>>>>> +  Add AV_FRAME_DATA_DOVI_METADATA.
>>>>> +
>>>>>  2021-12-xx - xxxxxxxxxx - lavf 59.10.100 - avformat.h
>>>>>  Add AVFormatContext io_close2 which returns an int
>>>>>
>>>>> diff --git a/libavutil/dovi_meta.c b/libavutil/dovi_meta.c
>>>>> index 7bd08f6c54..60b4cb2376 100644
>>>>> --- a/libavutil/dovi_meta.c
>>>>> +++ b/libavutil/dovi_meta.c
>>>>> @@ -33,3 +33,15 @@ AVDOVIDecoderConfigurationRecord *av_dovi_alloc(size_t *size)
>>>>>
>>>>>  return dovi;
>>>>>  }
>>>>> +
>>>>> +AVDOVIMetadata *av_dovi_metadata_alloc(size_t *size)
>>>>> +{
>>>>> +    AVDOVIMetadata *dovi = av_mallocz(sizeof(AVDOVIMetadata));
>>>>> +    if (!dovi)
>>>>> +        return NULL;
>>>>> +
>>>>> +    if (size)
>>>>> +        *size = sizeof(*dovi);
>>>>> +
>>>>> +    return dovi;
>>>>> +}
>>>>> diff --git a/libavutil/dovi_meta.h b/libavutil/dovi_meta.h
>>>>> index 299911d434..25e6d7b42f 100644
>>>>> --- a/libavutil/dovi_meta.h
>>>>> +++ b/libavutil/dovi_meta.h
>>>>> @@ -29,6 +29,7 @@
>>>>>
>>>>>  #include <stdint.h>
>>>>>  #include <stddef.h>
>>>>> +#include "rational.h"
>>>>>
>>>>>  /*
>>>>>  * DOVI configuration
>>>>> @@ -67,4 +68,146 @@ typedef struct AVDOVIDecoderConfigurationRecord {
>>>>>  */
>>>>>  AVDOVIDecoderConfigurationRecord *av_dovi_alloc(size_t *size);
>>>>>
>>>>> +/**
>>>>> + * Dolby Vision RPU data header.
>>>>> + *
>>>>> + * @note Cannot be extended without an ABI bump.
>>>>> + */
>>>>> +typedef struct AVDOVIRpuDataHeader {
>>>>> +    uint8_t rpu_type;
>>>>> +    uint16_t rpu_format;
>>>>> +    uint8_t vdr_rpu_profile;
>>>>> +    uint8_t vdr_rpu_level;
>>>>> +    uint8_t chroma_resampling_explicit_filter_flag;
>>>>> +    uint8_t coef_data_type; /* informative, lavc always converts to fixed */
>>>>> +    uint8_t coef_log2_denom;
>>>>> +    uint8_t vdr_rpu_normalized_idc;
>>>>> +    uint8_t bl_video_full_range_flag;
>>>>> +    uint8_t bl_bit_depth; /* [8, 16] */
>>>>> +    uint8_t el_bit_depth; /* [8, 16] */
>>>>> +    uint8_t vdr_bit_depth; /* [8, 16] */
>>>>> +    uint8_t spatial_resampling_filter_flag;
>>>>> +    uint8_t el_spatial_resampling_filter_flag;
>>>>> +    uint8_t disable_residual_flag;
>>>>> +} AVDOVIRpuDataHeader;
>>>>> +
>>>>> +enum AVDOVIMappingMethod {
>>>>> +    AV_DOVI_MAPPING_POLYNOMIAL = 0,
>>>>> +    AV_DOVI_MAPPING_MMR = 1,
>>>>> +};
>>>>> +
>>>>> +/**
>>>>> + * Coefficients of a piece-wise function. The pieces of the function span the
>>>>> + * value ranges between two adjacent pivot values.
>>>>> + *
>>>>> + * @note Cannot be extended without an ABI bump.
>>>>> + */
>>>>> +#define AV_DOVI_MAX_PIECES 8
>>>>> +typedef struct AVDOVIReshapingCurve {
>>>>> +    uint8_t num_pivots;                         /* [2, 9] */
>>>>> +    uint16_t pivots[AV_DOVI_MAX_PIECES + 1];    /* sorted ascending */
>>>>> +    enum AVDOVIMappingMethod mapping_idc[AV_DOVI_MAX_PIECES];
>>>>> +    /* AV_DOVI_MAPPING_POLYNOMIAL */
>>>>> +    uint8_t poly_order[AV_DOVI_MAX_PIECES];     /* [1, 2] */
>>>>> +    int64_t poly_coef[AV_DOVI_MAX_PIECES][3];   /* x^0, x^1, x^2 */
>>>>> +    /* AV_DOVI_MAPPING_MMR */
>>>>> +    uint8_t mmr_order[AV_DOVI_MAX_PIECES];      /* [1, 3] */
>>>>> +    int64_t mmr_constant[AV_DOVI_MAX_PIECES];
>>>>> +    int64_t mmr_coef[AV_DOVI_MAX_PIECES][3/* order - 1 */][7];
>>>>> +} AVDOVIReshapingCurve;
>>>>> +
>>>>> +enum AVDOVINLQMethod {
>>>>> +    AV_DOVI_NLQ_NONE = -1,
>>>>> +    AV_DOVI_NLQ_LINEAR_DZ = 0,
>>>>> +};
>>>>> +
>>>>> +/**
>>>>> + * Coefficients of the non-linear inverse quantization. For the interpretation
>>>>> + * of these, see ETSI GS CCM 001.
>>>>> + *
>>>>> + * @note Cannot be extended without an ABI bump.
>>>>> + */
>>>>> +typedef struct AVDOVINLQParams {
>>>>> +    uint64_t nlq_offset;
>>>>> +    uint64_t vdr_in_max;
>>>>> +    /* AV_DOVI_NLQ_LINEAR_DZ */
>>>>> +    uint64_t linear_deadzone_slope;
>>>>> +    uint64_t linear_deadzone_threshold;
>>>>> +} AVDOVINLQParams;
>>>>> +
>>>>> +/**
>>>>> + * Dolby Vision RPU data mapping parameters.
>>>>> + *
>>>>> + * @note Cannot be extended without an ABI bump.
>>>>> + */
>>>>> +typedef struct AVDOVIDataMapping {
>>>>> +    uint8_t vdr_rpu_id;
>>>>> +    uint8_t mapping_color_space;
>>>>> +    uint8_t mapping_chroma_format_idc;
>>>>> +    AVDOVIReshapingCurve curves[3]; /* per component */
>>>>> +
>>>>> +    /* Non-linear inverse quantization */
>>>>> +    enum AVDOVINLQMethod nlq_method_idc;
>>>>> +    uint32_t num_x_partitions;
>>>>> +    uint32_t num_y_partitions;
>>>>> +    AVDOVINLQParams nlq[3]; /* per component */
>>>>> +} AVDOVIDataMapping;
>>>>> +
>>>>> +typedef struct AVDOVIColorMetadata {
>>>>> +    uint8_t dm_metadata_id;
>>>>> +    uint8_t scene_refresh_flag;
>>>>> +
>>>>> +    /**
>>>>> +     * Coefficients of the custom Dolby Vision IPT-PQ matrices. These are to be
>>>>> +     * used instead of the matrices indicated by the frame's colorspace tags.
>>>>> +     * The output of rgb_to_lms_matrix is to be fed into a BT.2020 LMS->RGB
>>>>> +     * matrix based on a Hunt-Pointer-Estevez transform, but without any
>>>>> +     * crosstalk. (See the definition of the ICtCp colorspace for more
>>>>> +     * information.)
>>>>> +     */
>>>>> +    AVRational ycc_to_rgb_matrix[9]; /* before PQ linearization */
>>>>> +    AVRational ycc_to_rgb_offset[3]; /* input offset of neutral value */
>>>>> +    AVRational rgb_to_lms_matrix[9]; /* after PQ linearization */
>>>>> +
>>>>> +    /**
>>>>> +     * Extra signal metadata (see Dolby patents for more info).
>>>>> +     */
>>>>> +    uint16_t signal_eotf;
>>>>> +    uint16_t signal_eotf_param0;
>>>>> +    uint16_t signal_eotf_param1;
>>>>> +    uint32_t signal_eotf_param2;
>>>>> +    uint8_t signal_bit_depth;
>>>>> +    uint8_t signal_color_space;
>>>>> +    uint8_t signal_chroma_format;
>>>>> +    uint8_t signal_full_range_flag; /* [0, 3] */
>>>>> +    uint16_t source_min_pq;
>>>>> +    uint16_t source_max_pq;
>>>>> +    uint16_t source_diagonal;
>>>>> +} AVDOVIColorMetadata;
>>>>> +
>>>>> +/**
>>>>> + * Combined struct representing a combination of header, mapping and color
>>>>> + * metadata, for attaching to frames as side data.
>>>>> + *
>>>>> + * @note The struct must be allocated with av_dovi_metadata_alloc() and
>>>>> + *       its size is not a part of the public ABI.
>>>>> + */
>>>>> +
>>>>> +typedef struct AVDOVIMetadata {
>>>>> +    AVDOVIRpuDataHeader header;
>>>>> +    AVDOVIDataMapping mapping;
>>>>> +    AVDOVIColorMetadata color;
>>>>>
>>>>
>>>> I think we should version this by adding an integer version
>>>> field, and a compile-time header #define.
>>>> just in case more extra signalling is added. Then
>>>> we could support it in a sort of backwards-compatible way
>>>> by just documenting what has changed, and users could
>>>> make their own decisions about whether to present it
>>>> as-is or error out.
>>>>
>>>
>>> That seems rather pointless to me when API users can just tie that to
>>> avutil or avcodec version.
>>> Additionally, how would I ever make use of that? We never support
>>> running binaries older then the headers you have, and newer data can't
>>> be used because you don't have the struct definition that corresponds
>>> with it.
>>>
>>
>> One way to make these structs extensible while keeping this side data a
>> single buffer is by adding the offsets of header, mapping and color to
>> the start of AVDOVIMetadata at the start of AVDOVIMetadata. That way
>> users could use the parts they know and ignore the rest. It would even
>> allow for adding further structs to AVDOVIMetadata.
>>
> 
> How would you define the struct then? If we define it as a normal
> struct and people naively access stuff, it might suddenly point to
> something else - so if we allow us to change it, then we probably
> should define it in a way that doesn't even allow the naive accidental
> mis-reading.
> Keeping that in mind then, it sounds like it would get rather
> complicated to use.
> 

Similar to how it is done in detection_bbox.h: Add some inline accessor
functions. We could even remove the substructs from the publically
visibly part of the struct altogether and only allow access through
these accessors.

- Andreas


More information about the ffmpeg-devel mailing list