experiment(vegas): vegas_scroll.prefetch_gate runs the prefetch only while the render thread waits on vsync

The render thread spends most of each refresh in SwapOnVSync with the GIL
released, then needs it back the moment the swap returns. With plugin
rendering on the prefetch thread, it often has to wait for it -- behind
bytecode for up to the switch interval, behind a GIL-holding C call for as
long as that takes -- and hdpi's late frames of 2-5 refreshes went up.

src/common/render_gate.py opens a window around each swap, up to just
before the refresh the swap will return on, and a profile hook on the
prefetch thread parks it outside that window. It is never parked holding a
lock the render thread also takes (the Vegas buffer, cache and state
locks, logging, threading, importlib, the cache), never when no frame has
been swapped for 50ms, and never for more than 50ms at a time.

Off by default and ignored on a binding that keeps the GIL in
SwapOnVSync, where the window would never let the prefetch run.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

src/common/render_gate.py

@@ -0,0 +1,222 @@
+"""Let a background thread run Python only while the render thread waits on vsync.
+
+With plugin rendering moved to Vegas's prefetch thread (DisplayManager.offscreen,
+#630) the render thread no longer stops for it, but it still shares the GIL
+with it. The render thread spends most of each refresh inside SwapOnVSync,
+which releases the GIL, and needs it back the moment the swap returns. If the
+prefetch thread is running Python right then, the render thread waits: up to
+the switch interval (5ms) behind bytecode, and for as long as a C call that
+keeps the GIL takes. On hdpi that showed up as frames 2-5 refreshes late while
+a group was being prepared.
+
+The gate turns that around. The display manager opens it just before each swap,
+with a deadline shortly ahead of the refresh the swap will return on, and
+closes it when the swap returns. A thread inside ``gate.yielding()`` checks it on
+every Python and C call through a profile hook, and once the window has closed
+it parks -- blocked on a condition, GIL released -- until the next swap opens
+it. The render thread then finds the GIL free when its refresh arrives, and the
+background work runs in time the render thread was only spending waiting.
+
+Parking a thread is only safe if nothing the render thread needs is stuck
+behind it, so it is never parked:
+
+* while it holds a lock registered with ``guard()`` (the Vegas buffers and
+  caches the render thread also takes);
+* inside logging, threading, importlib or the cache, all of which take locks the
+  render thread can take too;
+* when there is no render loop to protect -- no swap for ``STALE_SECONDS``, as
+  on a static screen or a stalled frame.
+
+And a parked thread is never held more than ``MAX_WAIT_SECONDS`` at a time, so
+whatever the gate gets wrong costs a frame, not a freeze.
+
+Only worth enabling on a binding whose SwapOnVSync releases the GIL (see
+scripts/build_rgbmatrix_nogil.sh). With one that keeps it, the window never
+lets the background thread run and it only makes progress in the timeouts.
+"""
+
+from __future__ import annotations
+
+import math
+import sys
+import threading
+import time
+from collections import deque
+from typing import Any, Callable, Deque, List, Optional
+
+#: Park background threads this long before the refresh a swap will return on,
+#: so a short C call already under way has finished by then.
+MARGIN_SECONDS = 0.002
+
+#: The longest a background thread is parked in one go.
+MAX_WAIT_SECONDS = 0.05
+
+#: No swap for this long means there is no render loop running to protect.
+STALE_SECONDS = 0.05
+
+#: Swaps needed before the refresh period is trusted enough to open a window.
+MIN_SAMPLES = 8
+
+#: Parking with any of these on the stack could hold a lock the render
+#: thread takes: logging handler locks, Condition and Event internals, the
+#: module import locks, and the disk and memory cache locks.
+_UNSAFE_PATHS = (
+    "logging",
+    "threading.py",
+    "importlib",
+    "cache",
+)
+
+
+def _unsafe(frame: Any, base: Any) -> bool:
+    """True if a frame above ``base`` comes from somewhere parking could deadlock.
+
+    ``base`` is the frame that entered ``yielding()``; what lies below it (the
+    thread's own bootstrap in threading.py) holds nothing.
+    """
+    while frame is not None and frame is not base:
+        filename = frame.f_code.co_filename
+        for part in _UNSAFE_PATHS:
+            if part in filename:
+                return True
+        frame = frame.f_back
+    return False
+
+
+def swap_releases_gil() -> Optional[bool]:
+    """Whether the loaded rgbmatrix binding releases the GIL, or None if none is loaded.
+
+    The rebuilt binding links PyEval_SaveThread and the stock one never does.
+    The same test as src.common.frame_timing.binding_releases_gil (#629); one
+    of the two goes once both have landed.
+    """
+    module = sys.modules.get("rgbmatrix.core")
+    path = getattr(module, "__file__", None)
+    if not path:
+        return None
+    try:
+        with open(path, "rb") as handle:
+            return b"PyEval_SaveThread" in handle.read()
+    except OSError:
+        return None
+
+
+def _held(lock: Any) -> bool:
+    """Is ``lock`` held? RLocks report this thread's ownership; plain locks, anyone's."""
+    is_owned = getattr(lock, "_is_owned", None)
+    if is_owned is not None:
+        return is_owned()
+    return lock.locked()
+
+
+class RenderGate:
+    """Opened by the render thread around each swap; honoured by background threads."""
+
+    def __init__(self, clock: Callable[[], float] = time.monotonic):
+        self.clock = clock
+        self._cond = threading.Condition()
+        self._generation = 0
+        self._open_until = 0.0
+        self._last_return: Optional[float] = None
+        self._periods: Deque[float] = deque(maxlen=64)
+        self._period: Optional[float] = None
+        self._guarded: List[Any] = []
+        self._local = threading.local()
+        #: How often, and for how long in all, background threads were parked.
+        self.parks = 0
+        self.parked_seconds = 0.0
+
+    def guard(self, *locks: Any) -> None:
+        """Never park a thread while it holds (or, for a plain Lock, anyone holds) these."""
+        self._guarded.extend(lock for lock in locks if lock is not None)
+
+    # -- render thread -----------------------------------------------------
+
+    def refresh_period(self) -> Optional[float]:
+        """The panel's refresh period from recent swaps, or None until known.
+
+        The 10th percentile of the gaps between swap returns, each divided by
+        the hold: a late frame only ever lengthens a gap, so the low end is
+        the panel's own period.
+        """
+        return self._period
+
+    def before_swap(self, hold: int) -> None:
+        """The render thread is about to block in SwapOnVSync: open the window."""
+        hold = max(1, int(hold))
+        period = self._period
+        now = self.clock()
+        last = self._last_return
+        if period and last is not None and now - last < STALE_SECONDS:
+            # The swap returns on the first refresh boundary after both the
+            # current frame's hold is up and this frame has been handed over;
+            # boundaries fall a whole period apart from the last return.
+            refreshes = max(hold, math.ceil((now - last) / period))
+            open_until = last + refreshes * period - MARGIN_SECONDS
+        else:
+            open_until = 0.0     # no rhythm to predict from: leave threads be
+        with self._cond:
+            self._open_until = open_until
+            self._generation += 1
+            self._cond.notify_all()
+
+    def after_swap(self, hold: int) -> None:
+        """The swap returned and the render thread needs the GIL: close the window."""
+        now = self.clock()
+        self._open_until = 0.0
+        last = self._last_return
+        if last is not None and now - last < STALE_SECONDS:
+            self._periods.append((now - last) / max(1, int(hold)))
+            if len(self._periods) >= MIN_SAMPLES:
+                ordered = sorted(self._periods)
+                self._period = ordered[len(ordered) // 10]
+        self._last_return = now
+
+    # -- background threads ------------------------------------------------
+
+    def _should_park(self, frame: Any, now: float) -> bool:
+        if now < self._open_until:
+            return False                 # inside the window
+        last = self._last_return
+        if last is None or now - last > STALE_SECONDS or self._period is None:
+            return False                 # no render loop to protect
+        for lock in self._guarded:
+            if _held(lock):
+                return False
+        return not _unsafe(frame, getattr(self._local, "base", None))
+
+    def _hook(self, frame: Any, _event: str, _arg: Any) -> None:
+        now = self.clock()
+        if not self._should_park(frame, now):
+            return
+        generation = self._generation
+        with self._cond:
+            self._cond.wait_for(lambda: self._generation != generation,
+                                timeout=MAX_WAIT_SECONDS)
+        self.parks += 1
+        self.parked_seconds += self.clock() - now
+
+    def yielding(self) -> "_Yielding":
+        """``with gate.yielding():`` runs the block giving way to the render thread."""
+        return _Yielding(self)
+
+
+class _Yielding:
+    """Installs a gate's profile hook on the thread for the length of a block."""
+
+    def __init__(self, gate: RenderGate):
+        self.gate = gate
+        self._previous: Any = None
+        self._previous_base: Any = None
+
+    def __enter__(self) -> RenderGate:
+        local = self.gate._local  # pylint: disable=protected-access
+        self._previous_base = getattr(local, "base", None)
+        local.base = sys._getframe(1)  # pylint: disable=protected-access
+        self._previous = sys.getprofile()
+        sys.setprofile(self.gate._hook)  # pylint: disable=protected-access
+        return self.gate
+
+    def __exit__(self, *_exc: Any) -> None:
+        sys.setprofile(self._previous)
+        self.gate._local.base = self._previous_base  # pylint: disable=protected-access

src/display_manager.py

@@ -322,6 +322,11 @@ class DisplayManager:
         # See src/common/scroll_config.py and scripts/scroll_speeds.py.
         self._frame_hold = 1
 
+        # A src.common.render_gate.RenderGate while Vegas runs with
+        # vegas_scroll.prefetch_gate on: opened around each swap so the
+        # prefetch thread only runs Python while this thread waits on vsync.
+        self.render_gate = None
+
         self._scrolling_state = {
             'is_scrolling': False,
             'last_scroll_activity': 0,
@@ -948,7 +953,12 @@ class DisplayManager:
                 # Swap buffers immediately. framerate_fraction holds the frame
                 # for N refreshes; SwapOnVSync blocks for all of them, which is
                 # what paces the render loop to the chosen frame rate.
+                gate = self.render_gate
+                if gate is not None:
+                    gate.before_swap(self._frame_hold)
                 self.matrix.SwapOnVSync(self.offscreen_canvas, self._frame_hold)
+                if gate is not None:
+                    gate.after_swap(self._frame_hold)
 
                 # Swap our canvas references
                 self.offscreen_canvas, self.current_canvas = self.current_canvas, self.offscreen_canvas

src/vegas_mode/config.py

@@ -90,6 +90,14 @@ class VegasModeConfig:
     # Experimental: measured with scripts/frame_soak.py before it gets a default.
     switch_interval_ms: float = 0.0
 
+    # Let the prefetch thread run Python only while the render thread is
+    # blocked waiting for vsync, and park it the rest of the time, so the
+    # render thread never waits for the GIL when its refresh comes round. Needs
+    # a binding that releases the GIL in SwapOnVSync; ignored otherwise.
+    # Experimental: see src/common/render_gate.py and measure with
+    # scripts/frame_soak.py before it gets a default.
+    prefetch_gate: bool = False
+
     # Keep one continuous strip, extending it with the next group of plugins as
     # the scroll approaches the end, instead of composing a fresh strip and
     # swapping it in. A swap stops the motion, substitutes every pixel at once
@@ -221,6 +229,7 @@ class VegasModeConfig:
             continuous_scroll=vegas_config.get('continuous_scroll', True),
             offscreen_prefetch=bool(vegas_config.get('offscreen_prefetch', True)),
             switch_interval_ms=float(vegas_config.get('switch_interval_ms', 0.0) or 0.0),
+            prefetch_gate=bool(vegas_config.get('prefetch_gate', False)),
             extend_threshold_screens=float(
                 vegas_config.get('extend_threshold_screens', 2.0)),
             auto_trim=vegas_config.get('auto_trim', True),
@@ -264,6 +273,7 @@ class VegasModeConfig:
             'continuous_scroll': self.continuous_scroll,
             'offscreen_prefetch': self.offscreen_prefetch,
             'switch_interval_ms': self.switch_interval_ms,
+            'prefetch_gate': self.prefetch_gate,
             'extend_threshold_screens': self.extend_threshold_screens,
             'auto_trim': self.auto_trim,
             'trim_threshold': self.trim_threshold,

src/vegas_mode/coordinator.py

@@ -18,6 +18,7 @@ import time
 import threading
 from typing import Optional, Dict, Any, List, Callable, TYPE_CHECKING
 
+from src.common import render_gate
 from src.vegas_mode.config import VegasModeConfig
 from src.vegas_mode.plugin_adapter import PluginAdapter
 from src.vegas_mode.stream_manager import StreamManager
@@ -282,6 +283,7 @@ class VegasModeCoordinator:
             self._fps_last_health_log = 0.0
             self._fps_was_degraded = False
             self._apply_switch_interval()
+            self._install_render_gate()
 
         # Line up the next group immediately, so the first extension is already
         # warm rather than stalling the scroll to fetch it.
@@ -305,6 +307,7 @@ class VegasModeCoordinator:
                 self._start_time = None
 
         self._restore_switch_interval()
+        self._remove_render_gate()
 
         # Cleanup components
         self.render_pipeline.reset()
@@ -330,6 +333,38 @@ class VegasModeCoordinator:
             sys.setswitchinterval(saved)
             self._saved_switch_interval = None
 
+    def _install_render_gate(self) -> None:
+        """Gate the prefetch thread on the render thread's swaps; see VegasModeConfig."""
+        if not self.vegas_config.prefetch_gate:
+            return
+        if getattr(self.display_manager, 'render_gate', None) is not None:
+            return
+        releases = render_gate.swap_releases_gil()
+        if not releases:
+            logger.warning(
+                "Vegas: prefetch_gate ignored -- %s",
+                "this rgbmatrix binding keeps the GIL in SwapOnVSync "
+                "(scripts/build_rgbmatrix_nogil.sh)" if releases is False
+                else "no hardware binding loaded")
+            return
+        gate = render_gate.RenderGate()
+        # Locks the render thread takes too: never park the prefetch holding one.
+        gate.guard(self._state_lock,
+                   getattr(self.stream_manager, '_buffer_lock', None),
+                   getattr(self.render_pipeline, '_buffer_lock', None),
+                   getattr(self.render_pipeline, '_prefetch_lock', None),
+                   getattr(self.plugin_adapter, '_cache_lock', None))
+        self.display_manager.render_gate = gate
+        logger.info("Vegas: prefetch gated on vsync")
+
+    def _remove_render_gate(self) -> None:
+        gate = getattr(self.display_manager, 'render_gate', None)
+        if gate is None:
+            return
+        self.display_manager.render_gate = None
+        logger.info("Vegas: prefetch gate parked the prefetch %d times, %.1fs in all",
+                    gate.parks, gate.parked_seconds)
+
     def pause(self) -> None:
         """Pause Vegas mode (for live priority interruption)."""
         with self._state_lock:
@@ -552,6 +587,10 @@ class VegasModeCoordinator:
                         fps, target, fps_frame_count,
                         p99 * 1000.0, frame_worst * 1000.0
                     )
+                    gate = getattr(self.display_manager, 'render_gate', None)
+                    if gate is not None:
+                        logger.info("Vegas: prefetch parked %d times, %.1fs in all",
+                                    gate.parks, gate.parked_seconds)
                     self._fps_last_health_log = current_time
                 else:
                     logger.debug(

src/vegas_mode/render_pipeline.py

@@ -10,6 +10,7 @@ import os
 import time
 import threading
 from collections import deque
+from contextlib import nullcontext
 from typing import Optional, List, Any, Dict, Deque, TYPE_CHECKING
 from PIL import Image
 
@@ -392,8 +393,12 @@ class RenderPipeline:
                     os.nice(10)
                 except (OSError, AttributeError):
                     pass
+                # With vegas_scroll.prefetch_gate on, run only while the render
+                # thread waits on vsync; see src/common/render_gate.py.
+                gate = getattr(self.display_manager, 'render_gate', None)
                 try:
-                    group = self.stream_manager.take_next_group(offscreen_only=True)
+                    with gate.yielding() if gate is not None else nullcontext():
+                        group = self.stream_manager.take_next_group(offscreen_only=True)
                 except Exception:
                     logger.exception("Background prefetch failed")
                     group = []