fix: audio pump starvation on large files with audio (#1041)

Root causes fixed:
1. Remove starvation check that suspends AudioContext — it freezes the
   video clock and causes playback to stall permanently
2. Add backpressure (1s BUFFER_AHEAD) to prevent audio pump from decoding
   unbounded buffers, starving the video decoder
3. Add setTimeout(0) yield between batches so video rAF callbacks can run
4. Batch audio buffer reads (16 per iteration) to reduce IPC overhead
5. Fix latestScheduledEndTime time-domain: use media-time (same domain as
   currentTime) instead of AudioContext-time, so backpressure check works
   correctly after seeking

Without these fixes, playing a large file (9+ GB) with audio causes:
- Audio pump decodes hundreds of buffers in 1 second (195s of audio decoded
  in 1s wall clock)
- Main thread blocked by 177 IPC calls/sec from audio iterator
- Video frames never render (rAF starved)
- After seek: backpressure never triggers due to time-domain mismatch,
  causing persistent stutter and eventual silence

Author: iodone

packages/artplayer-proxy-mediabunny/src/AudioEngine.js

@@ -126,84 +126,90 @@ export default class AudioEngine {
     await this.stopIterator()
     this.audioIterator = this.audioSink.buffers(this.currentTime)
 
+    // Batch size: read multiple audio buffers per iteration to reduce
+    // IPC round-trip overhead. 16 buffers ≈ 340ms of audio.
+    const BATCH_SIZE = 16
+
     while (true) {
       if (localId !== this.asyncId || this.paused)
         return
 
-      const nextPromise = this.audioIterator.next()
-
-      // Monitor for buffer starvation
-      const checkStarvation = setInterval(() => {
-        if (localId !== this.asyncId || this.paused) {
-          clearInterval(checkStarvation)
-          return
-        }
-
-        if (
-          this.audioContext.state === 'running'
-          && this.audioContext.currentTime >= this.latestScheduledEndTime - 0.2
-        ) {
-          this.audioContext.suspend()
-          this.events.emit('waiting')
-        }
-      }, 50)
-
-      let result
+      const batch = []
+      let batchDone = false
       try {
-        result = await nextPromise
+        for (let i = 0; i < BATCH_SIZE; i++) {
+          const result = await this.audioIterator.next()
+          if (result.done) {
+            batchDone = true
+            break
+          }
+          batch.push(result.value)
+        }
       }
       catch (e) {
         console.error('Audio iterator error:', e)
-        break
-      }
-      finally {
-        clearInterval(checkStarvation)
+        batchDone = true
       }
 
       if (localId !== this.asyncId || this.paused)
         return
 
       // Resume if was suspended
-      if (this.audioContext.state === 'suspended') {
+      if (batch.length > 0 && this.audioContext.state === 'suspended') {
         await this.audioContext.resume()
         this.events.emit('canplay')
         this.events.emit('playing')
       }
 
-      if (result.done)
-        break
+      // Schedule all buffers in the batch
+      for (const { buffer, timestamp } of batch) {
+        const node = this.audioContext.createBufferSource()
+        node.buffer = buffer
+        node.connect(this.gainNode)
+        node.playbackRate.value = this.playbackRate
 
-      const { buffer, timestamp } = result.value
+        const startAt
+          = this.audioContextStartTime
+            + (timestamp - this.playbackTimeAtStart) / this.playbackRate
 
-      // Schedule audio buffer
-      const node = this.audioContext.createBufferSource()
-      node.buffer = buffer
-      node.connect(this.gainNode)
-      node.playbackRate.value = this.playbackRate
+        const duration = buffer.duration
+        const endAt = startAt + duration / this.playbackRate
 
-      const startAt
-        = this.audioContextStartTime
-          + (timestamp - this.playbackTimeAtStart) / this.playbackRate
+        const endMediaTime = (endAt - this.audioContextStartTime) * this.playbackRate + this.playbackTimeAtStart
+        if (endMediaTime > this.latestScheduledEndTime) {
+          this.latestScheduledEndTime = endMediaTime
+        }
 
-      const duration = buffer.duration
-      const endAt = startAt + duration / this.playbackRate
+        if (startAt >= this.audioContext.currentTime) {
+          node.start(startAt)
+        }
+        else {
+          node.start(
+            this.audioContext.currentTime,
+            (this.audioContext.currentTime - startAt) * this.playbackRate,
+          )
+        }
 
-      if (endAt > this.latestScheduledEndTime) {
-        this.latestScheduledEndTime = endAt
+        this.queuedNodes.add(node)
+        node.onended = () => this.queuedNodes.delete(node)
       }
 
-      if (startAt >= this.audioContext.currentTime) {
-        node.start(startAt)
-      }
-      else {
-        node.start(
-          this.audioContext.currentTime,
-          (this.audioContext.currentTime - startAt) * this.playbackRate,
-        )
-      }
+      if (batchDone)
+        break
+
+      // Yield main thread after each batch so video rAF callbacks can run.
+      // Without this, the audio pump's tight loop starves the render loop.
+      await new Promise(resolve => setTimeout(resolve, 0))
 
-      this.queuedNodes.add(node)
-      node.onended = () => this.queuedNodes.delete(node)
+      // Backpressure: if we've scheduled too far ahead, wait for playback
+      // to catch up before decoding more. This prevents the audio pump from
+      // consuming all WebCodecs resources and starving the video decoder.
+      const BUFFER_AHEAD = 1 // seconds
+      while (this.latestScheduledEndTime - this.currentTime > BUFFER_AHEAD) {
+        if (localId !== this.asyncId || this.paused)
+          return
+        await new Promise(resolve => setTimeout(resolve, 50))
+      }
     }
   }
 
@@ -220,7 +226,7 @@ export default class AudioEngine {
     }
 
     this.audioContextStartTime = this.audioContext.currentTime
-    this.latestScheduledEndTime = this.audioContextStartTime
+    this.latestScheduledEndTime = this.playbackTimeAtStart
     this.paused = false
 
     const id = ++this.asyncId
@@ -241,7 +247,7 @@ export default class AudioEngine {
   async seek(time) {
     this.playbackTimeAtStart = Math.max(0, time)
     this.audioContextStartTime = this.audioContext.currentTime
-    this.latestScheduledEndTime = this.audioContextStartTime
+    this.latestScheduledEndTime = this.playbackTimeAtStart
 
     const id = ++this.asyncId
     if (!this.paused) {