Symptom: properly-converted BT.2020 PQ/HLG clips rendered progressively
desaturated vs the SDR reference — the signature of the transfer/primaries tags
not reaching the NV12→RGB shader, so HDR code values were decoded as plain
sRGB/BT.709 (no gamut expansion, wrong EOTF).
VAAPI hardware-decoded frames frequently leave color_trc/color_primaries/
colorspace/color_range unspecified — the authoritative values live on the codec
context (parsed from the bitstream), which the importer never sees. Before
importing, copy any unspecified colour field from the codec context onto the
frame, so the importer detects PQ/HLG + BT.2020 correctly.
Also add a one-time importer log (LB_VIDEO_DEBUG) of the detected
ten_bit/full_range/colorspace/primaries/trc to confirm what's picked up.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Decode HDR video into the linear compositor correctly instead of approximating
everything as sRGB/BT.709:
- Read the frame's color_trc and color_primaries in the importer → VideoTransfer
{Gamma,Pq,Hlg} + VideoPrimaries {Bt709,Bt2020} on GpuVideoFrame.
- nv12_blit.wgsl: branch the EOTF — sRGB gamma (SDR), SMPTE2084 PQ (normalized so
203-nit graphics white = 1.0; highlights exceed 1.0), or HLG inverse-OETF
(reference white ≈ 1.0). Then BT.2020→BT.709 primaries in linear light when
wide-gamut, clamping out-of-709 colours.
Establishes the white=1.0 scene-linear convention: SDR content is unchanged
(stays in [0,1]); HDR video carries super-white highlights through compositing.
SDR-output mapping (clip default vs highlight rolloff) is Part 2. HLG's display
OOTF is omitted (scene-referred) — approximate but reasonable for SDR-out.
The NV12→RGB pass hardcoded BT.709, so SD (BT.601) clips had slightly wrong hues.
Read each frame's AVColorSpace in the importer and derive the Y'CbCr→R'G'B'
matrix (BT.709/601/240M/2020; Unspecified guessed by height like swscale/players),
carry the four coefficients on GpuVideoFrame, and apply them in the shader.
- core: GpuVideoFrame.coeffs + ycbcr_coeffs(kr, kb) helper.
- hw_video.rs: map AVColorSpace → (kr, kb) → coeffs.
- nv12_blit{.rs,.wgsl}: uniform grows to 80 bytes (adds a coeffs vec4); the matrix
multiply uses params.coeffs instead of literals.
BT.2020's transfer is still approximated as sRGB. The DRM-modifier-without-SAMPLED case stays a graceful software fallback.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Make the dormant core HW-decode engine live for the preview path:
- hw_video.rs: editor's HwVideoImporter — maps a decoded VAAPI surface to a
DRM-PRIME DMA-BUF and imports it as wgpu NV12 plane textures on the *shared*
device (the only one with the import extensions). install() creates the VAAPI
device and injects it + the importer into the VideoManager.
- main.rs: track whether the shared device is actually in use; only then (Linux,
not LB_NO_SHARED_DEVICE) install hardware decode, using the CreationContext's
shared device + adapter.
- nv12_blit.rs + nv12_blit.wgsl: NV12 plane textures → BT.709 → sRGB-encoded →
linear, written straight into the Rgba16Float HDR layer (no CPU upload). Colour
math mirrors the software path so HW/SW video match; honours full_range.
- stage.rs: the preview Video arm branches on inst.gpu (NV12 blit) vs rgba_data
(existing upload+blit_straight); sets render_hardware_ok = !cpu_renderer so the
CPU fallback still gets software frames.
- video_exporter.rs: sets render_hardware_ok(false) before both compositing
passes — export composites on the encoder's separate device, so a hardware
decoder downloads to CPU instead (export stays software, correct).
- dmabuf.rs: imported plane textures now also carry SAMPLED/TEXTURE_BINDING so
they can be sampled by the NV12 blit (they were render-target-only); into_planes
hands the textures to the longer-lived GpuVideoFrame.
- video.rs: cache-key the GPU/CPU representation on want_gpu (HW-configured AND
render_hardware_ok) so software-only decode keeps a single cache entry.
Preview only this pass; export GPU-residency is the 3c-export follow-up. Untested
at runtime here (no GPU/display in container) — both crates compile.
Extend the existing VideoDecoder with an optional hardware path, reusing its
demux/seek/keyframe/blob engine (no duplication):
- GpuVideoFrame (NV12 plane wgpu textures) + HwVideoImporter trait (editor
implements the DMA-BUF import; the AVFrame crosses as an opaque pointer so
core needn't reference the GPU crate) + HwDeviceHandle (opaque AVBufferRef).
- VideoManager::set_hardware_decode injects the VAAPI device + importer; each
decoder attaches hw_device_ctx + a get_format(VAAPI) callback before opening,
decodes into VAAPI surfaces, and imports them (no CPU copy).
- get_frame returns DecodedFrame::{Cpu,Gpu}; VideoFrame/VideoRenderInstance gain
an optional `gpu`. The frame cache budgets GPU frames as ~w*h*3/2 and keys on
whether the consumer wanted GPU.
- A hardware decoder serving a CPU consumer (export, render_hardware_ok=false)
downloads the surface via av_hwframe_transfer_data then swscales — so export
stays software/correct and only the preview goes GPU-resident. HW init or
import failure falls back to software per clip.
Dormant until the editor injects an importer (next): no importer => software,
unchanged.
The decoder's output size was frozen to the document size at import, and export
reused that decoder — so exporting above document res upscaled the video (real
source detail discarded) and a document resize never re-targeted the decode.
Decode size is now chosen per get_frame call: VideoDecoder::get_frame and
VideoManager::get_frame take a target (w, h), capped to native (never upscale),
with the swscale context and frame caches keyed on the output size so preview
(preview res) and an in-progress export (export res) don't collide. The renderer
derives the target from the document->output base_transform, so export decodes
at export res (full detail) and the canvas at preview res. Thumbnails/asset
library pass small targets.
get_frame rebuilt the RGBA swscale context on every decoded frame and printed
[Video Timing] lines unconditionally. A stream's frames share one input
format/size, so build the scaler once (keyed on format+dims, rebuilt only if
they change) and reuse it; gate the per-frame traces behind LB_VIDEO_DEBUG.
Cuts export wall time ~10% on a 1080p video clip (the scaler rebuild was the
bulk of the per-frame "scale" cost; it's now ~0ms). SwsContext is !Send, so the
cached scaler is wrapped in a SendScaler — sound because a VideoDecoder is only
ever touched under the VideoManager mutex (same invariant as its decoder/input).
Video was the only media type always kept external (VideoClip.file_path),
so a project with video wasn't self-contained. Now video packs into the
SQLite container under the same large-media policy as audio (pack < 2 GB
unless the user chose Reference), and both the frames and the embedded audio
track decode by streaming directly from the blob — no temp files.
- New crate ffmpeg-blob-io: an AVIOContext-over-Read+Seek shim (BlobInput)
that lets ffmpeg demux from an arbitrary byte source. Isolates all the
unsafe FFI + ffmpeg ABI coupling (version-pinned =8.0.0/=8.0.1). Manual
Drop teardown order; AVSEEK_SIZE restores the read position (FFmpeg assumes
a size query doesn't move it — required for MP4 moov-at-end).
- Schema/save/load: VideoClip.media_id; save_beam packs/references video as
MediaKind::Video (keyed by clip id); load resolves packed vs referenced and
reports missing sources. A packed clip points its linked video-audio pool
entry's media_id at the video row so the audio streams from the same blob.
- Frames: video.rs VideoSource{Path,Packed} threaded through new/seek/scan/
probe/thumbnails (a fresh BlobReader per open); editor builds the source
from current_file_path (now set before register_loaded_videos).
- Audio: VideoAudioReader::open_source via BlobInput; the disk_reader
StreamSource block on packed video-audio is removed; the engine's existing
factory activation routes it unchanged.
Tests: ffmpeg-blob-io AVIO unit tests (WAV via Cursor, seek, open/drop loop);
core packed_video_stream (blob->AVIO->Input) and beam_archive video round-trip;
daw-backend open_source test (compiles; links/runs only off-container).
Runtime-verified: a packed video plays frames + audio after the source file
is removed.
Thumbnail rendering fixes:
- Strip now tiles from each clip's true (unclamped) origin and draws only the
tiles intersecting the visible rect, so it scrolls correctly and shows the
right frames when a clip is scrolled partly off the left. Both render sites
(collapsed group + expanded track) share one draw_video_thumbnail_strip helper.
- On-clip strip no longer freezes on the first thumbnail: get_thumbnail_at now
returns the actual thumbnail timestamp and the GPU texture cache keys on it, so
tiles refresh as closer thumbnails finish generating.
- Hover preview derives content time from the clip's true origin too (matches the
strip when scrolled off-screen).
- insert_thumbnail keeps the cache sorted + deduped (fixes a latent unsorted
binary_search bug, and makes concurrent restore + resume race-safe).
Thumbnail persistence (mirrors waveform persistence):
- MediaKind::Thumbnail rows, keyed by thumbnail_media_id(clip_id) (clip id XOR a
sentinel). Each clip's thumbnails PNG-encoded into one opaque LBTN blob (editor
owns the format), snapshotted cheaply (Arc clones) and encoded off the UI thread.
- Save writes the packs (kept in place on re-save); load reads them into
LoadedProject.thumbnail_blobs; the editor decodes + inserts them on a background
thread, so reload shows thumbnails instantly with no re-decode (even if the
source video file is missing).
- Partial sets are persisted with a complete flag and RESUMED on load:
generate_keyframe_thumbnails takes a should_skip predicate so a save made
mid-generation continues from where it left off instead of redoing the work.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- VideoManager.frame_cache: unbounded HashMap (grew per distinct frame during
playback) -> LruCache evicted by a 256MB byte budget. Byte-budget rather than
frame count is robust across resolutions (a 4K frame is ~33MB vs ~2MB at
800x600). unload_video pops per-clip keys (LruCache has no retain).
- mux_video_and_audio: stream-merge the two inputs by PTS with one pending
packet per stream (O(1) memory) instead of collecting every packet into Vecs
first (O(duration)). Output is byte-identical.
- export AAC: sanitize the planar-f32 path (non-finite -> 0, finite clamped to
[-1,1]) like the integer paths, with a one-time warning. A stray NaN/Inf
render sample no longer fails the whole export.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Migrate the .beam container to SQLite and stream media from it instead of
decoding whole files into RAM on import/load.
Container & large files:
- SQLite .beam container (beam_archive) with in-place transactional saves and an
incremental BlobReader; supports both packed (chunked blobs) and referenced
(external path) media, with a user preference + first-import prompt for files
over the large-media threshold.
Audio streaming:
- Stream packed compressed audio on load via an inversion-of-control blob factory
(AudioBlobSourceFactory): daw-backend defines the trait, core implements it
over BlobReader, so the audio engine stays container-agnostic.
- Bulk-activate disk streaming for all loaded clips after SetProject.
- Sample-accurate compressed seek (SeekMode::Accurate; Coarse mislands on VBR).
Video:
- Video frames decoded/streamed on demand; thumbnails generated asynchronously
on a dedicated decoder so import/load never blocks the UI.
- The video's audio track is streamed on demand via an ffmpeg VideoAudioReader
as a separate editable AudioClip (no /tmp WAV extraction).
Waveform overview:
- Streaming min/max LOD pyramid (waveform_pyramid), bounded memory, configurable
floor B; serialized into the container and restored on load (or generated in
the background from the packed blob when absent), so no re-decode on reload.
- GPU min/max upload path; integer-LOD textureLoad fixes zoom-dependent wobble.
Skyler Lehmkuhl