craft-audio names its VMD decoder feature `vmd-decode`, and craft-video
names a codec's decoder `<codec>-decode` wherever the codec may also have
an encoder (mpeg2, mpeg4, vp8…). The video decoder now follows that:
`vmd-decode` gates the module and its VideoFrame conversion, and `vmd`
turns it on, so `vmd` keeps working and an encoder can be added later
without renaming.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
real_files_match_ffmpeg waited for ffmpeg without looking at its exit
status, so pictures that matched before ffmpeg failed still passed. It now
asserts success, as the generated-file oracle does.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`VideoFrame::try_from(&vmd::Decoder)` looks the picture's palette indices
up in the decoder's palette and gives 8-bit R, G, B planes (matrix 0). The
README and the frame module docs said every decoder's output converts by
moving its buffers; paletted video copies, as VMD keeps only indices.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Preserves codec algorithm from AniCraft1f3e603b7ff53a3da364e41ec4da8d43dc83aab0 and2b7571363b4b647266df758f95b6e397021cfd4d; neutral frame type and explicit caller limits adapt the shared API.
mjpeg::Decoder called craft_jpeg::decode with no Adobe transform, so a
packet with numeric component ids and APP14 transform 0 (RGB as stored)
was converted as YCbCr and came out in the wrong colours. The decoder
now reads the transform with craft_jpeg::adobe_transform, as the JPEG
image readers do. A regression test injects APP14 transforms 0 and 1;
it fails without the fix.
we_decode_what_ffmpeg_encodes_to_what_ffmpeg_decodes built its input
with ffmpeg's gradients source, which picks random colors on every run,
so the agreement it measured changed from run to run. On one CI run the
4:2:0 case scored 30.5 dB against the 38 dB floor. The test now writes
the file's own picture(67, 41, 3) as raw RGB and feeds it to ffmpeg
through rawvideo, for both the plain and the default-tables encodes.
Addresses the CodeRabbit review on #1692: odd-width raw RGB rows carry a pad byte,
and BI_BITFIELDS is only BGRX with the matching masks (new from_avi_bitfields).
* image-codecs: decode AVIF pixels with craft-video's pure-Rust AV1 decoder
New optional `avif-decode` feature (native targets) reads still AVIF
files: the HEIF/ISOBMFF item graph, grids, clap/irot/imir, alpha
auxiliary images (including premultiplied), 8/10/12-bit 4:0:0, 4:2:0,
4:2:2 and 4:4:4, nclx/ICC colour, Exif and XMP, and PQ/HLG to linear
F32 through cicp::linearize as the HEIF path does.
The container walk follows LightCraft #3017 by @emmertex. Every size is
checked against the caller's Limits before the AV1 decoder runs, the
decoder's memory budget is capped by max_alloc, and the decode is
wrapped in catch_unwind so a decoder panic becomes an error.
Fixtures are tiny ffmpeg/libaom files; the 4:2:0 planes match libdav1d
bit for bit. Adds a decode_avif fuzz target.
Part of storytold/craft-libs#778.
* image-codecs, video: address review on the AVIF decoder
- clap: width and height are unsigned fractions, the offsets have signed
numerators, so 0x80000000/0x40000000 is a width of 2, not -2.
- An item with clap, irot or imir more than once (repeated ipma records)
is refused before any canvas work, instead of rotating 261120 times.
- Matrix coefficients 10 (BT.2020 constant luminance) is refused rather
than converted with the non-constant-luminance equations.
- An alpha plane whose own transforms differ from the colour's is refused
rather than composed in the wrong place; none, or the colour's, still
decode.
- craft-video: Decoder::set_shown_budget lowers the per-chunk byte cap on
shown pictures (never above 1 GiB), and the AVIF decoder sets it to
max_alloc, so a payload of many shown pictures is refused before they
are built.
A frame-level wrapper over craft-jpeg: packets decode to gray or RGB
frames (Annex K Huffman tables put in for AVI MJPG packets that have
none) and frames encode to baseline JPEG packets. Checked against
ffmpeg as a black-box oracle in both directions.
Part of #1510.
Adds the ffv1 feature (RFC 9043): YCbCr (any 2^n subsampling, gray,
optional transparency) and RGB through the JPEG 2000 reversible colour
transform, 8 to 16 bits, any slice raster, slice and configuration
record CRCs. Not covered yet: the Golomb-Rice coder, versions 0 and 1,
inter frames.
Part of #1509
- media: the frame buffer comes from a helper that returns `None` for a
size the address space cannot hold; a test asks it for sizes past
`isize::MAX`, where `vec!` panicked.
- video: the real-file comparison takes the palette of a frame without
a rectangle, as the documented loop does, instead of failing on it.
- media: `read_frame` reserves the frame's bytes fallibly. On a 32-bit
target a frame of over 2 GiB inside a larger source made `vec!` panic
with a capacity overflow; it is now an error.
- video: `Decoder::decode_palette` takes the palette of a frame that
has no rectangle in the picture, so the documented frame loop no
longer drops it. ffmpeg applies the palette of a frame whose
rectangle is empty; the pictures after it now agree.
- media: correct round 1's frame placement. The picture's position is
not one value for a file: it starts at the screen's corner and a
frame with data that is as large as the picture and not at the corner
moves it for the frames that follow, as ffmpeg was seen to do.
`picture_rect` now takes the frame's index; `origin()` is gone.
- audio: unknown frame flags are refused for an empty frame too.
- video: count the unpack buffer in the picture budget.
- docs: which sound frames are bounded by what, which cut-short blocks
are errors, and why sound needs both the flag and a rate.
- tests: ffmpeg's placement cases, the size of the LZ stop, multi-block
16-bit stereo frames, and an LZ test that compared half its picture.
- media: place video frames by the first frame as large as the picture,
not by the header's position fields, as ffmpeg was seen to do; a
frame outside the picture has no rectangle.
- media: a file has sound only with the header flag and a sample rate.
- video: stop LZ unpacking at what a picture can use, so a frame's cost
no longer goes by a length the file states.
- audio: a flags 1 sound frame plays every whole block it holds.
- docs: how sound lines up with pictures, what the buffer count is, and
which cut-short blocks are errors.
- tests: the frame-count cap, the 0xFF tag near the end of LZ data,
palette components over 6 bits, the error conversion and the new
identifiers.
Sierra VMD (Video and Music Data) is the video format of Sierra and
Coktel Vision games of the 1990s. Written clean-room from the public
format description at https://wiki.multimedia.cx/index.php/VMD, with
ffmpeg as a black-box oracle.
- craft-media: `vmd` feature, a demuxer for the 816-byte header and the
table of contents; `Container::Vmd`, `CodecId::VmdVideo` and
`CodecId::VmdAudio`.
- craft-video: `vmd` feature, a decoder for 8-bit paletted pictures
(rendering methods 1 to 3, LZ packing, palette changes).
- craft-audio: `vmd-decode` feature, a decoder for 8-bit PCM and 16-bit
DPCM sound, mono and stereo.
- craft-media-testkit: a VMD fixture writer, since ffmpeg reads VMD but
cannot write it.
Tests compare the decoders with ffmpeg on generated files, and on real
files when CRAFT_MEDIA_VMD_SAMPLES names a directory of them. 16-bit,
24-bit and Indeo 3 pictures, IMA ADPCM sound and the later header
revisions are not decoded.
- flush ends the GOP for average bitrate too: the budget of the pictures
the GOP did not get is taken back, so flushed short GOPs no longer
raise the bitrate.
- A lone picture flushed before a second sets the step gets dts = pts - 1,
so decoding times still increase if encoding goes on.
- Presentation times whose decoding time does not fit in an i64 are
refused with BadTimestamp before any state changes.
- Tests: the skip range test uses an f_code that can carry its vector, and
the reconstruction count is checked before comparing pictures.