Skip to content

Fix #22, #64: correct byte stride alignment and disable embedded still streams - #2403

Open
lulumumu9999 wants to merge 301 commits into
olive-editor:masterfrom
lulumumu9999:fix_resolution_22_64
Open

lulumumu9999 wants to merge 301 commits into
olive-editor:masterfrom
lulumumu9999:fix_resolution_22_64

Conversation

@lulumumu9999

Copy link
Copy Markdown

Summary

This PR fixes two related issues caused by incorrect memory stride handling and stream probing.

Changes

  • app/codec/frame.cpp: Changed generate_linesize_bytes() from 32-pixel alignment to 4-byte alignment. The previous logic (width + 31) & ~31 caused row offset mismatches when external libraries (FFmpeg, OpenImageIO) used different padding strategies, leading to diagonal tearing and buffer overflows with non-standard image resolutions.

  • app/codec/ffmpeg/ffmpegdecoder.cpp:

    • Fixed Probe() to automatically disable embedded still image streams (e.g., DJI camera thumbnails) when a regular video stream is also present, preventing the thumbnail from appearing as a video track on the timeline.
    • Fixed YUV plane texture format in ProcessFrameIntoTexture() to use U8/U16 instead of INVALID, which could cause undefined behavior during GPU texture creation.
  • app/codec/oiio/oiiodecoder.cpp: Explicitly passed buffer_.linesize_bytes() to OIIO::read_image() as the destination row stride. Previously, OIIO assumed tightly packed data, causing row misalignment when Oak's frame buffer had 4-byte aligned padding.

Issues Fixed

Fixes #22
Fixes #64

Mike-Solar and others added 27 commits July 17, 2026 06:24
…on nodes)

- node_audio_test: PanNode/VolumeNode sample math, TimeInput, ValueNode
- node_distort_test: transform/crop/flip/cornerpin/mask/ripple/swirl/
  tile/wave inputs, shader code, value jobs, resolution autoscale math
- node_filter_keying_test: opacity/blur/dropshadow/mosaic/stroke jobs,
  chroma/color-difference key and despill value composition
- node_math_transition_test: TrigonometryNode ops, MergeNode,
  TransitionBlock offsets/progress/crossfade math, SubtitleBlock,
  traverser globals and table caching
…render tail)

- node_save_load_test: Node Save/Load round trips for values, keyframes,
  hints, caches, connections, positions
- node_polygon_folder_test: Folder child management, PolygonGenerator
  rasterization/gizmos, TextGeneratorV1/V2
- footage_probe_test: real FFmpeg/OIIO probing of demo.mp4/img.png,
  metadata cache, footage state transitions
- render_tail_test: DynamicRenderer, color-context shader plumbing,
  PreviewAutoCacher pause/clear paths, DiskManager edges,
  AudioPlaybackCache segment I/O

Also fixes AudioPlaybackCache::WritePartOfSampleBuffer, found by the new
tests: zero padding was written via QFile::write(const char*) which
treats the buffer as a NUL-terminated string, so no padding bytes were
ever written; the write length was also computed from the segment end
instead of the range end, making WriteSilence() spin forever on an
empty buffer
Node core:
- MathNode/TrigonometryNode combo strings realigned with Operation enums
- mathbase scalar/vector operand pick no longer uses bitwise type checks
- NodeSetPositionAndDependenciesRecursively moves dependencies again
- RemoveAllKeyframes undo actually restores keyframes
- NodeGroup GetInputName null-deref guard, passthrough ids use input id
- NodeValueTable::Has is an exact type match; tag fallback only for
  empty tags; kStrCombo/kPushButton get data type names
- delete_all_keyframes no longer loops forever on unparented keyframes;
  keyframe-load failures propagate; rational interpolation falls back to
  double; OpacityEffect no longer leaks its internal MathNode

Audio/footage:
- AudioVisualWaveform: GetSummaryFromTime underflow OOB read, TrimIn
  prepend length bookkeeping, OverwriteSums source channel indexing
- PanNode inserts the pan value into the sample job (keyframed pan
  works); OutputParamsChanged is emitted on device change; PortAudio
  device indices are validated before Pa_GetDeviceInfo
- Footage: AdjustTimeByLoopMode no longer hangs/UBs on degenerate
  lengths, GetStreamIndex bounds-checked, CheckFootage clears stale
  state on missing files, failed probes are not cached,
  FootageDescription::Load requires its own root element

Render/track:
- ViewerOutput pushes the tagged samples value; TrackList disconnects
  the track-height lambda; GetTrackFromReference validity check
- RenderManager dummy backend: null-initialized threads, guarded
  decoder-cache/timer paths; Renderer::Destroy releases color cache
  shaders/textures; unknown dynamic backends no longer alias to oakgl
- SharedMemoryRegion POSIX attach validates segment size; ReadMessage
  skips blank lines instead of failing; GC counter clamped;
  IsRenderingCustomRange implemented; TimeOffsetNode gets a true
  inverse OutputTimeAdjustment; zero-speed clips return the held frame

Plugin/nodes:
- OliveClip: stored default region of definition is honored, on-demand
  images are cached; OliveHost sets host identity properties and logs
  instead of showing modal dialogs offscreen; Plugin.h dead decls gone
- DespillNode guards graph-less use with Rec.709 fallback; description
  typos fixed (despill, swirl); mosaic applies when only one axis
  matches; Windows-only Project filename separator test fixed
- ChromaKey tolerance inputs are now spelled "upper/lower_tolerance_in";
  old projects with the misspelled "tolerence" IDs keep loading (values
  and connections) via a new Node::GetInputIDForLegacyID() hook applied
  in LoadInput and connection deserialization
- flip/ripple/swirl/tile/wave node IDs moved to the consistent
  org.olivevideoeditor.Olive.* domain; NodeFactory::CreateFromID maps
  the old org.oliveeditor.* IDs so old projects still resolve
- Regression tests: legacy ChromaKey IDs (values + connections) and
  legacy factory IDs
- html.cpp: rgba() colors parsed with setRedF/GreenF/BlueF (0-1) while the
  writer emits 0-255 integers, so semi-transparent text colors lost their
  RGB on round-trip; parse with integer setters instead
- CLIProgressDialog: percentage padding compared normalized progress
  (0.0-1.0) against 10/100, so padding was always fully applied; compute
  the percentage first
- TimelineUndoPointer BlockTrimCommand: remove_block_from_graph_ was
  never initialized (UB on redo)
- TimelineUndoGeneral TransitionRemoveCommand: track_ was never
  initialized; GetRelevantProject() could dereference it before redo()
- ProjectLoadTask::Run(): failure path deleted project_ without
  resetting it, leaving GetLoadedProject() dangling
- common: Html doc<->HTML conversion (17), OIIOUtils (6)
- cli: CLIProgressDialog rendering and CLITaskDialog (8)
- timeline: undo commands for pointer/ripple/split/track/workarea (43)
  and general commands (22)
- task: project import/load, import error dialog, cache tasks (13)
- codec: OIIO decoder/encoder, planar file device, FFmpeg encoder (15)
- ui: HumanStrings (6), icons (4), StyleManager (4)
- SequencePreset::Save() wrote element "interlacing_" while Load() read
  "interlacing", losing the interlacing mode on preset round-trips;
  Save() now writes "interlacing" and Load() accepts both for backward
  compatibility with existing preset files
- ExportFormatComboBox: current_ was never initialized (UB on GetFormat
  before first user selection)
- ExportSubtitlesTab::SetSidecarEnabled() called setEnabled instead of
  setChecked, so restored export params could never re-enable sidecar
  subtitles
- h264section.h: static const int constants were odr-use unsafe (link
  error when referenced); changed to static constexpr
- editing: speedduration, keyframeproperties, markerproperties, sequence
  presets/parameters, footage properties/relink, project properties
- misc: about, action search, autorecovery scan, config base, disk cache,
  text, progress, render cancel, task, color, key sequence editor,
  remaining preferences tabs
- export: format combo box, audio/video/subtitles tabs, H.264 sections,
  advanced video and save-preset dialogs
- FrameRateComboBox: RepopulateList() left current index at -1 on first
  fill, so GetFrameRate() returned 0/1 instead of the displayed first
  entry until something called SetFrameRate()
- HandMovableView: default_drag_mode_ was never initialized; an early
  Core::ToolChanged signal would setDragMode() with it (UB)
- slider: SliderBase/FloatSlider/IntegerSlider mapping, clamping,
  display transforms, drag math, tristate, label substitution
- layouts: FlowLayout wrap math, ColumnedGridLayout placement
- combos: all standard combos, node combo box, color label combo
- misc: menu, file/path fields, toolbar, color button/wheel widgets,
  bezier, resizable scrollbar, hand-movable view, node value tree,
  pixel sampler, collapse button, clickable/focusable labels
Any caller would hit a link error; there were none, so the bug was
latent until the new widget tests called it.
- curve/keyframe views: connections, selection, undo commands
- time ruler / playback controls: time<->scene math, buttons, seek
- project explorer: view model hierarchy, MIME drag&drop, rename undo,
  folder heuristics, toolbar signals
- node table/tree/param views, task view, history, multicam play queue,
  timeline selections, node view scene
- MainWindowLayoutInfo: toXml() iterated open_sequences_ for the
  <viewers> section and fromXml() never parsed it, so open footage
  viewers were lost on layout save/load
- PanelWidget never set QObject::objectName, so
  PanelManager::GetPanelWithName() always returned nullptr and every
  caller (layout restore, NodeView param panel lookup) was dead code
- PanelManager::DestroyInstance() left instance_ dangling (UAF on any
  later RegisterPanel), unlike the other singletons
- all 19 panel classes: construction, titles, context/signal wiring,
  save/load data round-trips
- MainWindowLayoutInfo XML round-trip, MainStatusBar, offscreen
  MainWindow construction with standard panels and menus
- RatioDialog parsing (decimal and : / ; separators), validation
format_, video_codec_, audio_codec_, subtitle_sidecar_fmt_ and
subtitles_codec_ were left indeterminate by the constructor; reading
them before the corresponding Enable* call was UB. Initialize to the
kFormatCount/kCodecCount invalid sentinels.
- RenderManager: GPU-side members (context_, decoder_cache_,
  shader_cache_, auto_cacher_, worker_pool_, decoder_clear_timer_) were
  left uninitialized when the configured graphics backend is unknown
  (e.g. dummy); ViewerWidget then dereferenced garbage and crashed.
  Initialize them at declaration
- CurveView::SelectKeyframesOfInput ignored its reference parameter and
  selected keyframes of every connected track; select only the
  requested track's keyframes
- SeekableWidget::SeekToScenePoint dereferenced GetViewerNode()
  unconditionally; skip the playhead update when no viewer is connected
- LoadOTIOTask: unknown root schema leaked the freshly allocated
  project_ (delete + reset; OTIO is not enabled in local builds so this
  file is compile-verified by inspection only)

Locked by new tests: RenderManagerDummyBackend,
TimeRuler.SeekToScenePointWithoutViewerIsNoOp,
CurveViewTest.SelectKeyframesOfInputSelectsOnlyRequestedTrack
Only compiled on macOS via transitive includes; breaks on Linux/MSVC.
The test forces GraphicsBackend=opengl and builds a full MainWindow
with QOpenGLWidget-based viewer panels. On headless platforms whose
QPA cannot provide a GL context (Ubuntu/Windows CI on offscreen), the
app's GL calls crash on a non-current context (SEGFAULT in CI on
both). Probe with QOpenGLContext+QOffscreenSurface and GTEST_SKIP when
GL is unusable; the test still runs on platforms with real GL.
ColorManager::Init() (run by every Project construction) dereferenced
GetDefaultConfig() unconditionally. Any Project created before
SetUpDefaultConfig() crashed inside OCIO getCanonicalName on a null
config — the Windows CI SEGFAULT, where the suite order runs a
Project-creating test first. Reproduced locally by running
MainWindowLayoutInfo.AccessorsStoreAndRetrieve as the first suite.
QIcon::addFile() differs between Qt versions in whether entries for
nonexistent files keep the icon non-null (Ubuntu CI failure). Assert
on what matters: no usable pixmap or size is produced.
DirectoryIsValid() ignored try_to_create_if_not_exists and always
called mkpath(). On the Windows CI runner (which may create dirs at
the drive root) PathWidget validation both created bogus directories
and never flagged invalid paths; on Linux the mkpath just failed.

- filefunctions: only mkpath when try_to_create_if_not_exists is set
- TaskProjectLoadTest: Project::set_filename() stores native
  separators on Windows, normalize before comparing paths
- DialogProjectProperties: use a nonexistent file inside a temp dir
  instead of a fixed /definitely/... path that prior tests may have
  partially created on a writable drive
Version is no longer hardcoded: at configure time CMake uses the tag
name when HEAD is exactly on a tag (leading "v" stripped), otherwise
the first 8 hex digits of the commit hash. When git is unavailable
(e.g. source tarball) it falls back to the contents of version.txt,
which is now the single place to bump the release version.

Also remove the temporary qDebug() flood in the audio playback path
(ViewerWidget::QueueNextAudioBuffer / ReceivedAudioBufferForPlayback,
AudioManager::PushToOutput).
- Linux CD builds no longer disable Vulkan: the AppImage, deb, rpm and
  Arch packages now build and ship liboakvulkan (the AppImage deploys
  it via linuxdeploy --library so libvulkan is bundled too, and the
  verify step checks both backend libraries); deb/rpm dependencies gain
  libvulkan1/vulkan-loader; the Arch PKGBUILD gains vulkan-headers
- macOS CD installs vulkan-loader and exports VULKAN_SDK so
  find_package(Vulkan) locates the Homebrew loader (previously the
  Vulkan backend silently never built there)
- ffmpeg_bridge now installs to the standard lib directory with
  /../lib RPATH instead of the non-standard ffmpeg_bridge/bin
  layout (verified: editor, worker, oakgl and oakvulkan all resolve it)
- build guides (EN/ZH) document the macOS Vulkan backend dependencies
  and the VULKAN_SDK variable
…yte stride alignment and disable embedded still streams
@lulumumu9999
lulumumu9999 marked this pull request as ready for review August 26, 2026 10:36
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Scrubbing timeline with playhead erratically will crash the program Olive occasionally crashes on launch

3 participants