fix(vsti): crackle fix P1-P3 — SHM 8-slot ring + drain-delta pump + AudioWorklet sink

- SharedMemoryIPC.h/main.cpp: render into ring slot (idx & ringMask), 8 slots
- shm.rs: ring fields + read_ring_audio; lib.rs pump drains FULL delta, not just latest
- bridgeAudioNode.js: AudioWorklet primary (bridge-audio-sink), ScriptProcessor fallback
- bridgeAudioWorklet.js: ring 64 blocks, underrun fade, overrun drop + crossfade
- layout tests: sizeof 27560 (8-align via u64)
This commit is contained in:
2026-08-16 20:43:50 +07:00
parent ed26e1706d
commit cead8c8c71
8 changed files with 235 additions and 29 deletions
+76 -14
View File
@@ -1,17 +1,21 @@
// app/static/js/services/bridgeAudioNode.js
// Audio sink: consumes 'bridge-audio' PCM frames (native bridge) and plays them
// through a ScriptProcessorNode into the WebAudio graph (track FX / master bus).
// into the WebAudio graph. Primary path: AudioWorklet (audio thread, no main
// thread jank). Fallback: ScriptProcessorNode when worklet unavailable.
// VSTi crackle fix (FIX_VSTI_AUDIO_CRACKLE.md P1/P2): worklet ring absorbs pump
// jitter, underrun fades to 0 (no click), overrun drops oldest + crossfade.
(function () {
// Bridge renders one 256-sample block per ~5.8ms; ScriptProcessor fires once
// per 4096 samples (~16 blocks @44.1k). The ring MUST cover one full callback
// period or we underrun (silence gaps) — 24 blocks ~= 139ms headroom.
var RING_DEPTH = 24;
var RING_DEPTH = 24; // fallback ring (blocks of 256 samples)
var WORKLET_MAX = 64; // blocks buffered while worklet loads (~371ms @48k)
var SP_BUFFER = 4096;
var _chunks = []; // [{l: Float32Array(256), r: Float32Array(256)}]
var _chunks = []; // fallback: [{l: Float32Array(256), r: Float32Array(256)}]
var _preBuffer = []; // pending blocks while worklet module loads
var _spn = null;
var _gainNode = null;
var _workletNode = null;
var _ctx = null;
var _initialized = false;
var _useWorklet = false;
var _dest = null; // destination gain node da noi (chong connect trung)
function _getCtx() {
@@ -22,11 +26,45 @@
return _ctx;
}
function _setupWorklet() {
if (!_ctx.audioWorklet || !_ctx.audioWorklet.addModule) {
console.warn('[BridgeAudioNode] AudioWorklet unavailable — ScriptProcessor fallback');
return false;
}
var url = '/static/js/services/bridgeAudioWorklet.js?v=202608152200';
_ctx.audioWorklet.addModule(url).then(function () {
if (!_initialized) return; // disconnect() happened while loading
if (_useWorklet) return; // already set up (init retry)
try {
_workletNode = new AudioWorkletNode(_ctx, 'bridge-audio-sink', {
numberOfInputs: 0,
numberOfOutputs: 1,
outputChannelCount: [2]
});
_workletNode.connect(_gainNode);
_useWorklet = true;
// drain pre-buffer into worklet (order preserved)
while (_preBuffer.length) {
var b = _preBuffer.shift();
_workletNode.port.postMessage({ type: 'audio', l: b.l, r: b.r });
}
console.log('[BridgeAudioNode] AudioWorklet active');
} catch (err) {
console.warn('[BridgeAudioNode] AudioWorkletNode failed — ScriptProcessor fallback', err);
_useWorklet = false;
}
}).catch(function (err) {
console.warn('[BridgeAudioNode] addModule failed — ScriptProcessor fallback', err);
});
return true; // pending — worklet may become active later
}
function init(audioCtx) {
if (_initialized) return;
_ctx = audioCtx || _getCtx();
_gainNode = _ctx.createGain();
_gainNode.gain.value = 1.0;
// ScriptProcessor created but NOT connected until we know worklet failed.
_spn = _ctx.createScriptProcessor(SP_BUFFER, 0, 2);
_spn.onaudioprocess = function (e) {
var L = e.outputBuffer.getChannelData(0);
@@ -34,8 +72,6 @@
var outLen = L.length;
L.fill(0); R.fill(0);
if (!_chunks.length) return;
// Drain as many queued blocks as fit in this callback (16 max) — filling
// only one block per callback previously played 1/16 of the audio.
var written = 0;
while (written < outLen && _chunks.length) {
var blk = _chunks[0];
@@ -46,36 +82,62 @@
if (n >= blk.l.length) {
_chunks.shift();
} else {
// Partial block (only possible if block size != 256): keep remainder.
blk.l = blk.l.subarray(n);
blk.r = blk.r.subarray(n);
}
}
};
_spn.connect(_gainNode);
if (!_setupWorklet()) {
// No worklet at all — go straight to ScriptProcessor.
_spn.connect(_gainNode);
_useWorklet = false;
while (_preBuffer.length) _chunks.push(_preBuffer.shift());
while (_chunks.length > RING_DEPTH) _chunks.shift();
}
_initialized = true;
console.log('[BridgeAudioNode] initialized (ring depth ' + RING_DEPTH + ')');
}
function onAudio(l, r) {
if (!_initialized) init();
_chunks.push({ l: new Float32Array(l), r: new Float32Array(r) });
// Drop oldest on overrun — keeps latency bounded (~139ms max).
var blk = { l: new Float32Array(l), r: new Float32Array(r) };
if (_useWorklet && _workletNode) {
_workletNode.port.postMessage({ type: 'audio', l: blk.l, r: blk.r });
return;
}
if (!_useWorklet && _ctx && _ctx.audioWorklet && _ctx.audioWorklet.addModule) {
// Worklet loading — buffer temporarily (bounded), avoid fallback churn.
_preBuffer.push(blk);
while (_preBuffer.length > WORKLET_MAX) _preBuffer.shift();
return;
}
_chunks.push(blk);
while (_chunks.length > RING_DEPTH) _chunks.shift();
}
function flush() { _chunks = []; }
function flush() {
_preBuffer = [];
_chunks = [];
if (_useWorklet && _workletNode) {
try { _workletNode.port.postMessage({ type: 'flush' }); } catch (e) {}
}
}
function getOutputNode() { return _gainNode; }
function disconnect() {
flush();
if (_workletNode) {
try { _workletNode.disconnect(); } catch (e) {}
try { _workletNode.port.close(); } catch (e) {}
}
if (_spn && _gainNode) {
try { _spn.disconnect(); } catch (e) {}
try { _gainNode.disconnect(); } catch (e) {}
}
// Reset state so a later init() can rebuild (AudioRoutingEngine re-connect).
_initialized = false;
_useWorklet = false;
_workletNode = null;
_spn = null;
_gainNode = null;
_ctx = null;
@@ -0,0 +1,85 @@
// app/static/js/services/bridgeAudioWorklet.js
// VSTi crackle fix: audio sink chay tren audio thread (khong jank main thread).
// - Ring block 256 mau, toi da 64 block (~371ms @48k) — hap thu jitter pump.
// - Underrun: fade ve 0 thay vi click (hard silence).
// - Resume sau silence: fade vao tu 0.
// - Overrun: drop block cu nhat + crossfade 64 mau tai diem drop (het click).
class BridgeAudioSinkProcessor extends AudioWorkletProcessor {
constructor() {
super();
this._ring = [];
this._MAX = 64;
this._gain = 1.0;
this._lastOutL = 0.0;
this._lastOutR = 0.0;
this._skipFade = 0; // >0: crossfade dang chay sau overrun drop
this._skipFromL = 0.0;
this._skipFromR = 0.0;
this.port.onmessage = (e) => {
const m = e.data;
if (!m) return;
if (m.type === 'audio') {
this._ring.push({ l: m.l, r: m.r });
while (this._ring.length > this._MAX) {
// overrun: bo block cu nhat; lan consume toi se crossfade ranh gioi
this._ring.shift();
this._skipFade = 64;
}
} else if (m.type === 'flush') {
this._ring.length = 0;
this._gain = 1.0;
this._skipFade = 0;
}
};
}
process(inputs, outputs) {
const out = outputs[0];
if (!out || out.length < 2) return true;
const L = out[0];
const R = out[1];
const n = L.length;
let w = 0;
while (w < n) {
if (this._ring.length) {
const blk = this._ring[0];
const avail = blk.l.length;
const take = Math.min(avail, n - w);
if (this._skipFade > 0) {
for (let i = 0; i < take; i++) {
const t = i / 64;
L[w + i] = (this._lastOutL * (1 - t) + blk.l[i] * t) * this._gain;
R[w + i] = (this._lastOutR * (1 - t) + blk.r[i] * t) * this._gain;
}
this._skipFade = Math.max(0, this._skipFade - take);
} else {
for (let i = 0; i < take; i++) {
L[w + i] = blk.l[i] * this._gain;
R[w + i] = blk.r[i] * this._gain;
}
}
this._lastOutL = L[w + take - 1];
this._lastOutR = R[w + take - 1];
if (this._gain < 1.0) this._gain = Math.min(1.0, this._gain + take / 256.0);
w += take;
if (take >= avail) {
this._ring.shift();
} else {
blk.l = blk.l.subarray(take);
blk.r = blk.r.subarray(take);
}
} else {
// underrun: fade ve 0 (khong click)
for (let i = w; i < n; i++) {
this._gain = Math.max(0, this._gain - 1.0 / 256.0);
L[i] = this._lastOutL * this._gain;
R[i] = this._lastOutR * this._gain;
}
this._lastOutL *= this._gain;
this._lastOutR *= this._gain;
w = n;
}
}
return true;
}
}
registerProcessor('bridge-audio-sink', BridgeAudioSinkProcessor);
+1 -1
View File
@@ -40,7 +40,7 @@
<script src="/static/js/services/soundfontStorage.js?v=202607271016"></script>
<script src="/static/js/services/soundfontPlayer.js?v=202608101800"></script>
<script src="/static/js/services/nativeBridgeService.js?v=202608112200"></script>
<script src="/static/js/services/bridgeAudioNode.js?v=202608121700"></script>
<script src="/static/js/services/bridgeAudioNode.js?v=202608152200"></script>
<script src="/static/js/services/unifiedMidiRouter.js?v=202608112200"></script>
<script src="/static/js/services/audioRoutingEngine.js?v=202608112200"></script>
<script src="/static/js/services/aiGateway.js?v=202608037200"></script>
+11
View File
@@ -62,6 +62,17 @@ struct SharedAudioBufferIPC {
// liveness on this, so a slow (but alive) audio loop no longer looks dead.
// Append-only field: existing offsets stay stable.
volatile uint32_t heartbeat;
// VSTi crackle fix: multi-slot audio ring (appended after heartbeat — all
// offsets above stay stable). The bridge renders block N into slot
// (bridgeWriteIndex & ringMask), then increments bridgeWriteIndex (publish).
// The reader drains EVERY slot between its last index and writeIndex, in
// order, so a delayed reader no longer loses a whole block to single-slot
// overwrite (the old 1-slot gap). ringMask = slot count - 1 (power of 2);
// 8 slots ~= 42ms of audio @ 48k — covers a pump/WebView jank window.
volatile uint32_t ringMask;
float ringLeft[8][AUDIO_BLOCK_SIZE];
float ringRight[8][AUDIO_BLOCK_SIZE];
};
// G4.3: per-plugin sandbox SHM — double-buffered audio handoff.
+6 -1
View File
@@ -654,6 +654,8 @@ int main(int argc, char* argv[]) {
#else
(void)shmName; // POSIX shm mapping (shm_open) added when porting off Windows
auto* shmIPC = (SharedAudioBufferIPC*)std::calloc(1, sizeof(SharedAudioBufferIPC));
// VSTi crackle fix: 8-slot audio ring (see SharedMemoryIPC.h).
shmIPC->ringMask = 7;
if (!shmIPC) return 1;
#endif
@@ -754,9 +756,12 @@ int main(int argc, char* argv[]) {
};
// Render only [from, to) of the block — used by sample-accurate splitting.
// VSTi crackle fix: target the multi-slot ring slot for the block being
// built, so old blocks stay readable until the reader drains them.
auto renderSegment = [&](uint32_t from, uint32_t to) {
if (to <= from) return;
instruments.renderAll(shmIPC->masterLeft + from, shmIPC->masterRight + from, to - from);
uint32_t slot = shmIPC->bridgeWriteIndex & shmIPC->ringMask;
instruments.renderAll(shmIPC->ringLeft[slot] + from, shmIPC->ringRight[slot] + from, to - from);
};
double blockDurationMs = (double)block / sampleRate * 1000.0;
+12 -1
View File
@@ -65,7 +65,18 @@ int main() {
// 5. Layout stability — MUST match Rust src-tauri/src/shm.rs
assert(sizeof(SharedAudioBufferIPC::MidiEventIPC) == 12);
assert(sizeof(SharedAudioBufferIPC::ControlEventIPC) == 1040);
assert(sizeof(SharedAudioBufferIPC) == 11168); // G3.1: +4 heartbeat
// G4.5 ring: 11168 (G3.1 +heartbeat) + ringMask(4) + ringLeft(8192) + ringRight(8192)
// = 27556, rounded up to 8-byte struct alignment (u64 blockTimestamp) => 27560.
// Matches Rust size_of (src-tauri/src/shm.rs layout test).
assert(sizeof(SharedAudioBufferIPC) == 27560);
// 6. Audio ring round-trip (multi-slot, VSTi crackle fix)
ipc->ringMask = 7;
for (uint32_t i = 0; i < AUDIO_BLOCK_SIZE; ++i) ipc->ringLeft[3][i] = (float)i * 2.0f;
ipc->ringRight[3][0] = 42.0f;
assert(ipc->ringLeft[3][255] == 510.0f);
assert(ipc->ringRight[3][0] == 42.0f);
assert((ipc->bridgeWriteIndex & ipc->ringMask) < 8);
std::printf("SHM self-check OK (sizeof struct = %zu bytes)\n", sizeof(SharedAudioBufferIPC));
shm_close(h);
+22 -11
View File
@@ -451,23 +451,34 @@ pub fn run() {
restart_attempts = 0;
}
// Audio emission = writeIndex advancing (audio loop alive).
let audio_advanced = guard.as_ref().map(|shm| {
let idx = shm.write_index();
// VSTi crackle fix: drain the FULL delta from the multi-slot
// ring (8 slots) — a delayed pump no longer drops whole
// blocks that a single-slot SHM would have overwritten.
let mut audio_advanced = None;
if let Some(s) = guard.as_ref() {
let idx = s.write_index();
if idx != last_index {
let delta = idx.wrapping_sub(last_index);
// Only the last 8 blocks are still intact in the ring;
// if the pump stalled longer, older blocks are lost.
let start = if delta as usize > 8 { idx - 8 } else { last_index };
for k in start..idx {
let (l, r) = s.read_ring_audio(k);
let _ = pump_handle.emit(
"bridge-audio",
AudioFrame { l: l.to_vec(), r: r.to_vec() },
);
}
if delta as usize > 8 {
eprintln!("[tauri] audio pump lost {} blocks (stalled > ring)", delta - 8);
}
last_index = idx;
true
} else {
false
audio_advanced = Some(true);
}
});
}
if audio_advanced == Some(true) {
last_audio_change = std::time::Instant::now();
stall_logged = false;
let (l, r) = guard.as_ref().map(|s| s.read_audio()).unwrap_or(([0f32; shm::AUDIO_BLOCK_SIZE], [0f32; shm::AUDIO_BLOCK_SIZE]));
let _ = pump_handle.emit(
"bridge-audio",
AudioFrame { l: l.to_vec(), r: r.to_vec() },
);
}
// G3.1 FIX: nhả guard TRƯỚC stall log / watchdog kill+spawn.
// kill_bridge -> flush_bridge_state khóa chính ShmState mutex
+22 -1
View File
@@ -67,6 +67,13 @@ pub struct SharedAudioBufferIPC {
// independent of the audio loop. Watchdog liveness keys on this.
// Append-only field: offsets above are unchanged.
pub heartbeat: u32,
// VSTi crackle fix: multi-slot audio ring (appended after heartbeat —
// offsets above are unchanged). Bridge renders block N into slot
// (bridge_write_index & ring_mask), then increments bridge_write_index.
// Pump drains EVERY slot between its last index and write_index, in order.
pub ring_mask: u32,
pub ring_left: [[f32; AUDIO_BLOCK_SIZE]; 8],
pub ring_right: [[f32; AUDIO_BLOCK_SIZE]; 8],
}
pub struct Shm {
@@ -159,6 +166,17 @@ impl Shm {
(ipc.master_left, ipc.master_right)
}
/// Read ring slot for block index `idx` (slot = idx & ring_mask).
pub fn read_ring_audio(&self, idx: u32) -> ([f32; AUDIO_BLOCK_SIZE], [f32; AUDIO_BLOCK_SIZE]) {
let ipc = self.ipc();
let slot = (idx & ipc.ring_mask) as usize;
(ipc.ring_left[slot], ipc.ring_right[slot])
}
pub fn ring_mask(&self) -> u32 {
self.ipc().ring_mask
}
pub fn write_index(&self) -> u32 {
self.ipc().bridge_write_index
}
@@ -196,6 +214,9 @@ mod tests {
// native_bridge/tests/shm_selfcheck.cpp asserts the same sizes.
assert_eq!(std::mem::size_of::<MidiEventIPC>(), 12);
assert_eq!(std::mem::size_of::<ControlEventIPC>(), 1040);
assert_eq!(std::mem::size_of::<SharedAudioBufferIPC>(), 11168);
// 11168 + ring_mask(4) + ring_left(8*256*4) + ring_right(8*256*4)
// = 27556, rounded up to 8-byte struct alignment (u64 block_timestamp)
// => 27560. Matches MSVC sizeof (verified against the C header).
assert_eq!(std::mem::size_of::<SharedAudioBufferIPC>(), 27560);
}
}