ee40a02766d411704887a7e7123cd17c1e3720fe
7
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 vicrodh
|
931cb06d38 |
feat(qt): ease the visualizers on the pulse, and make Plasma react
TWO FINDINGS, both of which falsified something that was being assumed.
FLUIDITY IS NOT FRAME RATE. Driven at QBZ_PULSE_MS=10 the shell presented ~100
frames/s — qt-batches confirmed it, and the script now prints the pulse it ran
with so two logs cannot be confused — and the bars looked exactly as they had at
30. They would: every published frame was applied as a STEP, so the extra
presents repainted the same discrete values three times over. What the Slint
build does differently is not rate, it is `animate bar-h { 90ms }` — it moves
BETWEEN values.
So the interpolation is back, ON THE PULSE, and it costs nothing. The old
interpolator was expensive because of its private 16 ms Timer, not because it
interpolated: it presented at its own rate, unsynchronised with everything else.
Easing on the pulse presents exactly as often as stepping on the pulse did — one
repaint per period either way, only the value differs. Self-parking survives: a
converged ease writes nothing and schedules no repaint. The price is lag, and it
is the honest one — k = 0.45 is ~83% there in three pulses (~100 ms), matching
the reference's 90 ms, against 46 ms of FFT window and 77 ms of Rust smoothing
already in the chain.
The from/to/progress triple is gone with it, and its consumers moved to `cur`.
Note for the next reader: `SpectrumBand` drove all three of its modes from
`Connections` on those properties, and a Connections handler naming a property
that no longer exists is NOT a compile error — QML warns at runtime and simply
never fires. That is how the whole band went dead at once; grep for
`on<Name>Changed` when removing a property, not only for direct reads.
PLASMA WAS DRIVEN BY THE SLOWEST SIGNALS IN THE PIPELINE while the fast ones
travelled beside it unconnected. `transientAmp`, `bandsLo` and `bandsHi` were
published, parsed and stored, and never handed to the item — the shader declared
them and received zeros — so everything ran off the energy bands, which decay
x0.85 per frame (~473 ms) against the bars' x0.65 (~178 ms).
Worse, `beat` stops being a beat on dense material: it decays x0.88 per tick, so
above ~3 hits/s it never falls far enough to come back as a spike and settles
into a shallow ripple around 0.7. A splat multiplied by it becomes a steady
glow, and the busier the music the flatter the picture — the most frenetic input
produced the least motion. `beatAc` subtracts a slow density floor, so a hit
reads as a hit at any tempo.
The scene itself was rebuilt around the reference the owner filmed (Butterchurn
on a percussion-heavy track). Its recipe is three parts: one thin annulus
injected per frame that the feedback warp marches inward, so the concentric
rings ARE the last N frames of audio laid out in space; retention high enough
for that to stack (this held 0.85, where a trail died in ~15 frames and nothing
ever accumulated); and angular structure from reading the spectrum BY ANGLE.
But a CLOSED ring was wrong, and the owner named why: a perfect inner
circumference reads as a drawn shape, hard-edges the field, and at this
retention successive rings pack into one solid disc — which is what killed the
ambient feel. The annulus is gated by angle now, so a hit paints arcs the
conveyor carries inward, the circle never closes, and the gaps stay open for the
background. Noise bends, frays and thins those arcs so they are made of plasma
rather than of compass.
The worms are back at full strength with tighter cores and a contrast curve — a
raw gaussian is all shoulder and reads as haze — and their orbits now reach past
the arc band instead of being trapped inside it. Colour comes from ONE
album-derived triad, so everything used to be a shade of the same hue; a
Rodrigues hue turn splits it into worms, arcs and an ambient wash that sit apart
whatever album is playing. That wash is what stops the background being void for
the foreground to be a cut-out against. A soft limiter replaces the hard clamp:
at 0.945 retention a clamp pins the middle to white and the structure vanishes.
Release build clean, offscreen smoke green, 421/421.
cc
|
||
|
vicrodh
|
f46c5c95d4 |
perf(qt): one repaint pulse for the whole shell (GPU 95% -> 25%)
Qt Quick has no partial redraws: any dirty item presents the WHOLE window, and on the owner's hybrid laptop each full-window present costs ~1.2% GPU flat (KWin composite + dGPU scanout; area-independent, while the app's own render is 0.08 ms/frame). The bill was presents/s, and the shell ran unsynchronised ~30 Hz clocks plus two leaks: - QbzShell.pulseMs: ONE Rust-driven 33 ms edge that every continuous animator ticks on (knob QBZ_PULSE_MS). VizSettle now stashes on publish and applies on the pulse (its 16 ms interpolation timer is gone); AmbientField and ImmersiveAtmosphere ride the same edge. - VizSettle writes NOTHING while not live: the always-mounted but invisible immersive panels were dirtying 28 bar rects per FFT publish (named via QT_LOGGING_RULES=qt.quick.dirty=true). - AmbientField renders inline again: ShaderEffectSource scheduleUpdate() double-presents (minimal repro fbo_repro.qml in the investigation folder). - Background mode 2 (ImmersiveAtmosphere) composites its four drifting layers + base + dim + scrim in ONE opaque ShaderEffect (atmosphere.frag, baked by build.rs); numeric parity vs the Image stack verified (atmo-shader-parity-check.py, delta 0.0). The four-Image stack stays as the software/failure fallback. Measured on the owner's machine: 93-97% -> 25-26% GPU, ~34 W -> 12-13 W, ~33 presents/s in both background modes. Full story: qbz-nix-docs/qt-frontend/2026-08-11-scenegraph-batches/GPU-COST-INVESTIGATION.md km |
||
|
vicrodh
|
273325f151 |
revert(qt): restore the 16ms settle driver — frequency lever backfires
Measured on the Mac (samples + QSG_RENDER_TIMING, NPB Large, ambient
off): F0 62fps/155ms-s render / ~2.5ms-frame -> F1 51fps/246ms-s /
~4.8ms -> F1b 29fps/~277ms-s / ~9.5ms. Every frequency cut made each
frame MORE expensive faster than the rate dropped — per-frame cost on
Apple Silicon is dominated by wake/scheduling/power-state overhead
(parking the render thread between frames), not by scene work. The
best measured state is the ORIGINAL 16ms driver, so restore it (this
reverts
|
||
|
vicrodh
|
f2be1ccee8 |
perf(qt): stage visualizer publishes, commit them in the tick (macOS F1b)
QSG_RENDER_TIMING on the F1 build (Mac, NPB Large, ambient off, 3.5 min of playback) showed the F1 throttle only bought 62 -> ~51 fps, not 30: the producer publishes at ~30 Hz AND each publish wrote from/to/ progress directly — at()'s only dependencies — so every 33 ms period invalidated the window TWICE (publish + tick). ~51 fps x ~3.6 ms render kept the render thread at ~20-25% of a core. Stage the publish into pendFrom/pendTo (which nothing visual binds to) and commit + advance in the driver tick: ONE visual write per period. Pixel-identical output (the bars hold the settled old frame until the tick, exactly what progress=0 showed), ~30 fps instead of ~51. Audits green (resolution + singleton xref), build green. km |
||
|
vicrodh
|
9db2176d2e |
perf(qt): throttle the visualizer settle driver to 33 ms (macOS F1)
The 2026-08-02 macOS measurement (2026-08-02-macos-perf-fix-plan.md) named the actual burn: not the backend (real Metal, DPR 1, GPU at 50-57%) but per-tick invalidation. The VizSettle driver at 16 ms invalidated the scene ~62x/s; Qt Quick has no partial damage, so each tick re-rendered the whole 3440x1440 window (~2 Metal render passes), ~6x pricier per frame on the Metal path than on Linux/Mesa. Paused: 0%. The producer publishes at TARGET_FPS = 30 and the Rust EMA band-limits the signal to ~6 Hz, so 33 ms keeps one interpolated step per published frame — the render bill halves and the motion difference is at the edge of perception. Self-parking, the live gate and the Rust side are untouched. Expected on the Mac: ~30% -> ~15-18% CPU at NPB Large playing. km |
||
|
vicrodh
|
d4667f5ea9 |
perf(qt): rounded covers use a real mask, the spectrum ticks at 62Hz, the derivative cache is bounded
The track's "ShaderEffect / ColorOverlay / MultiEffect render NOTHING on the
software path this port runs on" doctrine was measured FALSE on 2026-07-29:
qbz-qt runs on the GPU (OpenGL RHI, threaded render loop, vsync 5.56ms). The
claim came from an offscreen session, and QT_QPA_PLATFORM=offscreen forces
software by definition. Every design it justified needed re-deciding.
- RoundedImage was a Canvas (renderTarget: Canvas.Image) that rasterised with
QPainter on the CPU AND uploaded a texture per repaint, for every cover in
the app, purely to round a corner identical in every frame of that image's
life. It is now layer.enabled + MultiEffect { maskSource } — an FBO rendered
once and cached for static content. Rectangle { radius; clip: true } cannot
do this: clip in Qt Quick is a RECTANGULAR scissor and ignores radius.
maskThresholdMin/maskSpreadAtMin are 0.5/1.0, NOT the 0.0/0.0 that looks
natural: 0.0 collapses Qt's mask smoothstep to step(0, alpha) and every
pixel with any alpha snaps fully opaque — measured 0 antialiased pixels at
radius 40 against the Canvas's 292. At 0.5/1.0 it is 308, ramp [0,51,219,255].
A software machine keeps the Canvas arm (GraphicsInfo.api), and
QBZ_QT_ROUND_MODE=canvas pins it on a GPU machine where the mask misbehaves.
- VizSettle drove the spectrum from FrameAnimation, i.e. every rendered frame:
180fps on this box for data the producer publishes at TARGET_FPS = 30. Now a
16ms Timer; the wall-clock maths is unchanged.
- SpectrumBand's per-bar y was a SECOND JS binding derived from height. An
anchor does the same arithmetic in C++ with no binding at all, and each FFT
frame is materialised into a plain array once per publish instead of being
read element by element across the JS/C++ boundary. Bindings per tick: bars
28 (was 56), waveform 48 (was 96), energy 5 (was 10).
QQuickPaintedItem was investigated and rejected with evidence: cxx-qt 0.7.3
emits X(QObject* parent) : base(parent), which does not compile against
QQuickPaintedItem(QQuickItem*), paint() is pure virtual, and there is zero
precedent for a custom base class in this workspace. Qt 6 ShaderEffect takes
no GLSL, only precompiled .qsb, which would make the build depend on a Qt
tool whose path differs on macOS.
- PlaylistCollage point-sampled its mosaic in its own Canvas
(Context2D.drawImage does not filter when it scales — measured RMSE 0.000
against nearest-neighbour). It now draws through RoundedImage.
- The scaled-derivative cache had no eviction at all and the .jpg -> .png key
change had orphaned 1251 files. Boot housekeeping sweeps the orphans once and
bounds the directory at 64 MB, LRU by mtime. Verified by effect: first run
deleted 1259 files / 10.05 MB, every run since reports 0, and the directory
is 14 MB. The eviction is only survivable because RoundedImage clears
_scaled on Image.Error and re-requests.
- For You artwork patched the rails in place instead of republishing the
document, which was resetting contentY. HomeView is destroyed on every
navigation, so the patch map is re-handed on mount through
QbzHome.refreshForyouArt() — otherwise the Radio rail and the Spotlight hero
went blank for the rest of the session after one navigation, and the skeleton
pulse Timer stayed pinned ON, defeating the fix it belonged to.
- Spotlight's placeholder gated on the local path instead of the remote url, so
it flashed a glyph on a cold cache where the Slint does not.
The retired doctrine is corrected wherever it was load-bearing for a decision.
The pre-baked icon tints are deliberately NOT reopened and no drop-shadows are
added: both are visual changes, not perf fixes.
cc
|
||
|
vicrodh
|
4b34b70f70 |
perf(qbz-qt): interpolate the spectrum band instead of raising the FFT rate
The band moved at the producer's 30 Hz. Raising TARGET_FPS would have been the obvious knob, but it lives in qbz-audio and is shared with Slint, and the amplitudes are already EMA-smoothed in Rust to roughly 6 Hz — a faster producer would have bought CPU and no smoothness. The published frames are now replayed across the render frames that follow them, which is what Slint's dock gets from its per-bar 90ms animate. A late publish degrades to a hold instead of a jump, and the motion runs at the display's rate rather than the data's. Splits the strip out of the dock into SpectrumBand, so the dock file owns only the gating and the layout arithmetic. cc |