R7-A (Qwen 3.8 Max): oakengine/display.h rewritten to the POD contract
from r7-pure-abi-plan.md - oak_video_params everywhere, opaque
texture/frame handles with retain/free protocol (engine-heap control
blocks), OakSharedBuffer refcounted wrapper for the QVariant playback
path. All TexturePtr/FramePtr gone from app (47 sites).
R7-B (Qwen 3.8 Max): liboakengine.so exports 3486 -> 19 C++ symbols
(version script oakengine.ver: oakengine_* plus the documented
oakgl/oakvulkan dlopen plugin ABI). oakengine-obj OBJECT library feeds
both the shared lib and the test binaries (-rdynamic so dlopen'd
backends resolve engine objects).
Fix (Kimi K3): producer/consumer type mismatch - viewerdisplay
unpacks OakSharedBufferPtr but viewer.cpp pushed raw void* handles,
so no frame ever reached the display widget (all 5 vulkan viewer
tests timed out with 'never received a texture'). Producers now wrap
with oak_make_shared_frame / oak_make_shared_texture(retain).
Verified: build 0 errors, ctest 45/45, nm U _ZN5olive = 0 in
oak-editor/oak-render-worker/oak-cli, 19 exported C++ symbols in
liboakengine.so (all documented plugin ABI).
Every app module now reaches liboakengine exclusively through
oakengine_* C calls, EngineEventBridge subscriptions and app-side
handle headers (cliphandle/keyframehandle/nodevaluehandle/oakvaluehelper).
Direct C++ command construction, engine signal connect()s, and engine
type usage in MOC-visible signatures are gone: 557 -> 0 undefined
olive:: symbols in oak-editor.
- playback timer uses the audio output device as its master clock: the
PortAudio callback counts consumed frames (including underrun
zero-fill) so video cannot drift away from what is heard; wall clock
remains as fallback when no clocked output is running
- output clock compensates for device output latency; new Preferences >
Audio buffer size setting (0 = auto)
- SampleBuffer::speed() now uses linear interpolation instead of
nearest-neighbor sampling
- regression tests: audio-clock driven timer (fwd/rev/speed), wall
clock fallback, interpolation correctness
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).
Replace every direct FFmpeg call in the editor and the render worker
with the pure C ffmpeg_bridge API, wrapped in thin C++ adapters that
preserve the original interfaces:
- avframeptr.h: olive::AVFrame adapter around FBFrame handles
- ffmpegutils: format conversion helpers on FB_* constants; the int
overload is renamed GetCompatibleBridgePixelFormat to avoid a silent
overload-resolution trap with the PixelFormat enum
- ffmpegdecoder/ffmpegencoder: rewritten as handle-based adapters over
FBDecoder/FBProbe/FBEncoder/FBScaler/FBResampler
- audioprocessor: FBAudioGraph push/pull adapter
- pluginrenderer/OliveClip: sws/pixdesc usage converted to FBScaler and
fb_pix_fmt_* queries
- AudioParams/channel layouts are plain uint64_t masks everywhere
Build integration: the root project no longer links FFMPEG directly;
only ffmpeg_bridge does. Binaries resolve the bridge library at runtime
via @loader_path inside the macOS app bundle (copied there post-build)
and via $ORIGIN/../ffmpeg_bridge/bin on Linux; on Windows the DLL is
installed next to the executables, so packages on all three platforms
ship the bridge library.
ffmpeg_bridge gains the extra API the adapters need:
fb_frame_make_writable, fb_decoder_get_format_duration,
fb_resampler_convert_frame, FB_PIX_FMT_YUV440P, SRT validation in
fb_probe_read_subtitle_stream, packed/planar fixes in
fb_encoder_write_audio, and a component-size fix in
fb_pix_fmt_component_size.
- Add FixChannelLayout() helper to AudioProcessor::Open() to fall back to a
default native channel layout when the input/output layout is unspecified,
custom, or has a zero mask. This prevents FFmpeg's abuffer/aformat filters
from rejecting 'channel_layouts=0x0' on Linux.
- Log audio processor open parameters, viewer audio queue state, and
AudioManager::PushToOutput device/stream status to help diagnose silent
playback on Linux.
- The playback backup timer was using timebase_dbl() in seconds as its
QTimer interval, which truncated to 0 ms for normal frame rates. This
caused the timer to fire continuously and starve the event loop when
the video display was hidden, e.g. in Show Waveform Only mode, freezing
the UI and stopping the playhead. Multiply by 1000 to use milliseconds.
- If PlayInternal has neither video nor audio to prequeue, call
FinishPlayPreprocess immediately so the backup timer starts and the
playhead advances even in waveform-only mode with no audio.
- AudioWaveformView now uses the video frame rate as its timebase
(falling back to the default sequence frame rate), preventing the
enormous scene rect that froze the UI when it previously used the
audio sample rate.
- Call UpdateSceneRect() after changing the waveform timebase so the
QGraphicsScene bounds are recomputed immediately.
- Refresh the waveform view's viewer/timebase whenever the waveform
visibility mode changes, ensuring the correct timebase is applied
even if the node was connected earlier with a different setting.
This commit fixes crashes (SIGSEGV in CImg blur_bilateral) and black-frame
corruption artifacts during playback and scrubbing on macOS Apple Silicon.
Root cause analysis:
1. macOS uses Tile-Based Deferred Rendering (TBDR). glFlush() does not
guarantee tile memory writeback, causing glReadPixels to read incomplete
tiles (black/corrupted frames) and cache them to disk.
2. Olive uses multiple shared OpenGL contexts (RenderProcessor contexts vs.
thread-local PluginRenderer context). glFinish() only waits for the
current context, not the shared context that produced the texture. CPU
readback in PluginRenderer could read partially-rendered tiles.
3. OlivePluginInstance and OliveClipInstance are not thread-safe. Concurrent
RenderProcessors could corrupt internal QMap/images_ and params_ via
setInputTexture/renderAction races.
Fixes:
- OpenGLRenderer::Flush() on macOS now uses glFinish() unconditionally.
- OpenGLRenderer::DownloadFromTexture() and OpenGLRenderer::Blit() insert
glFinish() before readback/detach to ensure tile writeback completes.
- PluginRenderer::RenderPlugin() now acquires a per-instance mutex to
serialize concurrent OFX render calls.
- Before CPU readback in PluginRenderer, flush the renderer that originally
produced each input texture, ensuring cross-context synchronization.
- RenderProcessor::ProcessVideoFootage() flushes after BlitColorManaged.
- Add black-frame detection in ProcessVideoCacheJob() to auto-purge TBDR-
corrupted cache files.
- Add diagnostic qDebug() logging in viewer, decoder, renderer, and plugin
paths to aid future debugging.
This commit removes redundant readbacks and uploads in the OpenFX
plugin pipeline, achieving zero-copy rendering for GL-capable
plugins and reducing CPU path overhead.
PluginRenderer (OpenGL path):
- Skip ReadbackTextureToFrame + ConvertFrameIfNeeded + Upload after
plugin render. The destination texture is already valid on GPU.
- Pass readback_cpu=false to setInputTexture() so input textures are
provided via loadTexture() (GL texture IDs) instead of being
downloaded to CPU Image buffers.
- Remove duplicate ReadbackTextureToFrame block between
getClipPreferences and the second setInputTexture call.
OliveClipInstance:
- Add optional bool readback_cpu=true to setInputTexture(). When
false, only params and input_textures_ are updated; CPU readback
and memcpy into Image are skipped.
- Add pruneImagesCache() to prevent unbounded growth of images_.
Input clips are limited to 8 cached frames; output clips are
left untouched.
Micro-optimizations:
- Replace per-row memcpy loops with single block memcpy when
src/dst strides are contiguous (common case in Olive pipeline).
- Remove dead GL_PREAMBLE macro definition.
fix(viewer): resolve playback head lag, frozen frames, and pause delay
Three playback pipeline behavioral issues are fixed:
1. Prequeue blocked playhead start:
Reduce kVideoPlaybackInterval from 0.5s to 0.1s. This lowers
prequeue length from 15–30 frames to 3–6 frames, so the
playback timer starts much sooner after pressing Play.
2. Frozen display during playback:
Relax the hard frame-drop logic in RendererGeneratedFrameForQueue.
When the queue has fewer than 2 frames, keep late frames instead
of dropping them, preventing the viewer from freezing entirely
when rendering cannot keep up with playback speed.
3. Delayed frame update after pause:
Cancel in-flight render tickets in PauseInternal() before
deleting queue watchers. Previously the render thread continued
processing stale playback frames, blocking the single-frame
render requested by UpdateTextureFromNode().
Refactor FFmpeg frame processing and improve focus management in panels. Add memory sanitization for debug builds and update KDDockWidgets integration.
Mike-Solar
Mike Solar