Skip to content

PhotoStore 0.20.5 — Concurrent 4K playback and compact audio

Choose a tag to compare

@akiyamasho akiyamasho released this 01 Nov 09:08
621e240

Video audio now occupies a thin waveform band inside each 48-point video row, with per-source mute. Standalone audio remains one compact editable row. A bounded texture cache reuses waveforms during playback.

Concurrent playback removes full-frame CMS/tile copies, copies whole tiles into the native shared mapped uploads, fuses eligible two-to-eight-source RGBA8 Normal stacks into one FP32 pass per chunk, and uses a host-sized worker pool under the existing shared retained-memory budget and eight-session ceiling. Eligible Apple Silicon video stacks address tile-major RGBA directly from shared GPU buffers; unsupported stacks and devices retain the texture path. Eligible Apple Silicon opaque H.264 4:2:0 previews use FFmpeg VideoToolbox decoding with a bounded software fallback at the next exact timestamp. Decoded output still crosses a bounded RGB24 pipe into CPU tiles; this is not zero-copy decode to Metal. Large movie surfaces avoid redundant exact bounds scans. Opaque 8-bit H.264 carries RGB24 through a smaller pipe and reconstructs RGBA tiles directly; other codecs retain RGBA. Video pre-roll waits for the starting frame's canvas completion before starting audio and anchoring the clock; stop/seek/rebase replace readiness signals.

On Apple M1 Max / 32 GiB, eight distinct original 3840×2160 SDR H.264/30 fps streams completed at 31.04 timeline frames/s through PhotoStore's hardware wgpu compositor. A paced 180-frame/30 fps run had 1.600 ms p95 completion lateness and 5 deadlines over 5 ms. Every source yielded 180 distinct sampled hashes. Authored opacity/visibility and history/clean state were preserved; final sampled GPU pixels matched CPU reference, and the full decoded reference frame retained identical encoded tiles. No resolution downgrade.

These are headless original-resolution GPU-completion measurements, excluding native window scanout/input/sound-device output. They do not establish universal CapCut parity. Native acceptance remains manual; other codecs/effects/precision or more sources may perform differently. Accounting is not a total process/codec/GPU memory cap. photostore-cli playback-bench exposes the scoped measurement.

Exact binary source: 621e2401b1035b922b66576e9a6425fc6ca9b067. One completed local normal app/CLI build; same verified artifacts are published and immediately atomically installed. Running documents/session are preserved; restart manually when convenient to load the new code. ARM64, locally signed with the persistent PhotoStore certificate; not Developer ID/not notarized. Scheduler stays disabled/unloaded with no pending replay. Publication triggers no build.

Implementation, evidence and limits · Architecture plan · GPT Image design/provenance