perf(scroll): extend and trim the Vegas strip in place

Every strip extension rebuilt the whole strip (np.concatenate: 2-2.6 ms for
a 10-14k px strip at 512x64 on a Pi 4) and every trim copied what was left
(1.2-1.8 ms), on the render thread. With ~4 ms of slack per refresh, every
extension frame on hdpi missed its refresh (5/5 in each soak run).

The strip now lives in a buffer with spare room; cached_array is a view of
its live columns. An append writes only the new columns (~0.2 ms), a trim
only moves the view's start, and the one full copy happens when the buffer
is reallocated (STRIP_SPARE_FACTOR 3: about once every two strip-lengths
scrolled). A cached_array set from outside -- the multi-display follower's
read-only one, create_scrolling_image's -- is never written through, and a
new strip lets the old buffer go. last_copy_bytes says what was copied, and
the Vegas frame-timing attribution reports that instead of the whole strip.

test_scroll_helper_in_place.py: the buffer is reused and only new columns
copied, trims copy nothing, reallocation when the room runs out, outside
arrays untouched, and random appends/trims/patches/scrolling checked frame
by frame against the old copying strip (mutation-checked).

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

CHANGELOG.md

@@ -418,6 +418,15 @@ read any of them:
 
 ### Scrolling
 
+- A Vegas strip extension no longer costs a late frame. Appending the next
+  group rebuilt the whole strip (`np.concatenate`, 2-2.6ms for a 10-14k px
+  strip at 512x64 on a Pi 4) and trimming copied what was left (1.2-1.8ms),
+  so on hdpi every extension frame missed its refresh. The strip now lives in
+  a buffer with spare room (`ScrollHelper.STRIP_SPARE_FACTOR`): an append
+  writes only the new columns (~0.2ms), a trim only moves the start, and the
+  one full copy happens when the buffer is reallocated, about once every two
+  strip-lengths scrolled. A strip set from outside (the multi-display
+  follower's) is never written through.
 - A Vegas strip extension costs the render thread about a third of what it
   did. Appending the next group and trimming what has scrolled past each
   rebuilt the strip's PIL image from its numpy array in full

src/common/scroll_helper.py

@@ -114,6 +114,18 @@ class ScrollHelper:
         self.cached_image = None   # see the property below
         self.cached_array: Optional[np.ndarray] = None  # Numpy array cache for fast operations
         self.total_scroll_width = 0
+        # An extended strip lives in a buffer with spare room after it, and
+        # cached_array is a view of the buffer's live columns: an append writes
+        # only the new columns, and a trim only moves the view's start. See
+        # append_content. _strip_view is the view this helper last made; a
+        # cached_array that is anything else was set from outside and is not
+        # written through.
+        self._strip_buffer: Optional[np.ndarray] = None
+        self._strip_view: Optional[np.ndarray] = None
+        self._strip_start = 0
+        #: Bytes the last append_content / drop_scrolled_prefix copied, for
+        #: frame-timing attribution (src/common/frame_timing.py note_op).
+        self.last_copy_bytes = 0
         
         # Pre-allocated buffer for output frame (reused to avoid allocations)
         self._frame_buffer: Optional[np.ndarray] = None
@@ -249,6 +261,7 @@ class ScrollHelper:
             self.total_scroll_width = 0
             self.cached_image = Image.new('RGB', (self.display_width, self.display_height), (0, 0, 0))
             self.cached_array = np.array(self.cached_image)
+            self._forget_strip_buffer()
             self.scroll_position = 0.0
             self.total_distance_scrolled = 0.0
             self.scroll_complete = False
@@ -284,6 +297,7 @@ class ScrollHelper:
         self.cached_image = full_image
         # Convert to numpy array for fast operations
         self.cached_array = np.array(full_image)
+        self._forget_strip_buffer()
         actual_image_width = full_image.width
         self.total_scroll_width = actual_image_width
         
@@ -712,12 +726,12 @@ class ScrollHelper:
             addition.paste(img, (x, 0))
             x += img.width + element_gap
 
-        # numpy concatenate, and no conversion back: the strip can be tens of
-        # thousands of columns wide and this runs on the render path. The PIL
-        # image is built from the array only if something reads it (see the
-        # cached_image property).
-        self.cached_array = np.concatenate(
-            (self.cached_array, np.array(addition)), axis=1)
+        # Written into the spare room after the strip, when there is some:
+        # the strip can be tens of thousands of columns wide and this runs on
+        # the render thread, where copying all of it (2-3 ms at 512x64 on a
+        # Pi 4) cost the frame after every extension. The PIL image is built
+        # from the array only if something reads it (see cached_image).
+        self.cached_array = self._extended_strip(np.asarray(addition))
         self._defer_image()
         self.total_scroll_width = self.cached_array.shape[1]
         self.scroll_complete = False
@@ -729,6 +743,42 @@ class ScrollHelper:
         )
         return True
 
+    #: Room an extended strip's buffer is given, as a multiple of what it
+    #: holds when (re)allocated. Trims free columns at the front and appends
+    #: use them at the back, so with 3x the buffer is reallocated -- the one
+    #: full copy -- about once every two strip-lengths scrolled.
+    STRIP_SPARE_FACTOR = 3.0
+
+    def _extended_strip(self, addition: np.ndarray) -> np.ndarray:
+        """The strip with ``addition`` after it, written in place when it fits."""
+        live = self.cached_array
+        assert live is not None
+        width = live.shape[1]
+        added = addition.shape[1]
+        buffer = self._strip_buffer
+        if (live is self._strip_view and buffer is not None
+                and self._strip_start + width + added <= buffer.shape[1]):
+            end = self._strip_start + width
+            buffer[:, end:end + added] = addition
+            self.last_copy_bytes = addition.nbytes
+        else:
+            total = width + added
+            buffer = np.empty((live.shape[0], max(total + 1, int(total * self.STRIP_SPARE_FACTOR)))
+                              + live.shape[2:], dtype=live.dtype)
+            buffer[:, :width] = live
+            buffer[:, width:total] = addition
+            self._strip_buffer = buffer
+            self._strip_start = 0
+            self.last_copy_bytes = live.nbytes + addition.nbytes
+        self._strip_view = buffer[:, self._strip_start:self._strip_start + width + added]
+        return self._strip_view
+
+    def _forget_strip_buffer(self) -> None:
+        """A new strip replaces the extended one: let its buffer go."""
+        self._strip_buffer = None
+        self._strip_view = None
+        self._strip_start = 0
+
     def drop_scrolled_prefix(self, keep_before: int = 0) -> int:
         """
         Discard columns that have already scrolled past, to bound memory.
@@ -766,9 +816,18 @@ class ScrollHelper:
         if cut <= 0:
             return 0
 
-        # .copy() so the original buffer is released rather than kept alive by
-        # a numpy view. The PIL image is deferred, as in append_content.
-        self.cached_array = self.cached_array[:, cut:].copy()
+        if self.cached_array is self._strip_view:
+            # Only the view's start moves; the columns behind it are reused
+            # when the buffer is next reallocated (append_content).
+            self._strip_view = self.cached_array[:, cut:]
+            self._strip_start += cut
+            self.cached_array = self._strip_view
+            self.last_copy_bytes = 0
+        else:
+            # Not a strip this helper extended: .copy() so the original
+            # buffer is released rather than kept alive by a view.
+            self.cached_array = self.cached_array[:, cut:].copy()
+            self.last_copy_bytes = self.cached_array.nbytes
         self._defer_image()
         self.total_scroll_width = self.cached_array.shape[1]
         self.scroll_position -= cut
@@ -881,6 +940,7 @@ class ScrollHelper:
         
         # Convert to numpy array for fast operations (required for get_visible_portion)
         self.cached_array = np.array(image)
+        self._forget_strip_buffer()
         
         # Update scroll width
         self.total_scroll_width = image.width
@@ -1138,6 +1198,7 @@ class ScrollHelper:
         """
         self.cached_image = None
         self.cached_array = None
+        self._forget_strip_buffer()
         self.total_scroll_width = 0
         self.scroll_position = 0.0
         self.total_distance_scrolled = 0.0

src/vegas_mode/render_pipeline.py

@@ -466,6 +466,11 @@ class RenderPipeline:
         if note is not None:
             note(kind, nbytes)
 
+    def _copied_bytes(self) -> int:
+        """Bytes the scroll helper's last append or trim copied (the strip, if it cannot say)."""
+        copied = getattr(self.scroll_helper, 'last_copy_bytes', None)
+        return int(copied) if isinstance(copied, int) else self._strip_nbytes()
+
     def _strip_nbytes(self) -> int:
         array = self.scroll_helper.cached_array
         return int(array.nbytes) if array is not None else 0
@@ -660,7 +665,7 @@ class RenderPipeline:
             element_gap=0,
         )
         if appended:
-            self._note_op('extend', self._strip_nbytes())
+            self._note_op('extend', self._copied_bytes())
             logger.info(
                 "[%s] Appended deferred content: strip now %dpx, %dpx ahead",
                 plugin_id, self.scroll_helper.total_scroll_width,
@@ -770,7 +775,7 @@ class RenderPipeline:
             )
             if not appended:
                 return False
-            moved = self._strip_nbytes()
+            moved = self._copied_bytes()
 
             # Where each block starts, laid out as append_content does: a
             # separator before every block, or -- when there was no strip to
@@ -789,9 +794,10 @@ class RenderPipeline:
                 self._static_markers = tuple(
                     (max(0, x - cut), pid) for x, pid in self._static_markers)
             self._forget_trimmed_records(cut)
-            # The append built the whole strip anew, and a trim copies what is
-            # left of it again: both land in the frame after this one.
-            self._note_op('extend', moved + (self._strip_nbytes() if cut else 0))
+            # Both land in the frame after this one: the append's new columns
+            # (the whole strip when its buffer had to be reallocated), and a
+            # trim's copy, if it made one.
+            self._note_op('extend', moved + (self._copied_bytes() if cut else 0))
 
             self._segments_in_scroll = [pid for pid, _ in grouped]
             self.stats['composition_count'] += 1

test/test_scroll_helper_in_place.py

@@ -0,0 +1,146 @@
+"""ScrollHelper extends and trims a strip in place (src/common/scroll_helper.py).
+
+Every extension of the Vegas strip used to rebuild it whole (np.concatenate)
+and every trim copied what was left: 3.5-4.5 ms on the render thread at 512x64
+on a Pi 4, so the frame after each extension was late. The strip now lives in
+a buffer with spare room: an append writes only the new columns, a trim only
+moves the view's start, and a full copy happens only when the buffer is
+reallocated. What a frame shows must not change at all.
+"""
+import random
+import sys
+from pathlib import Path
+
+import numpy as np
+import pytest
+from PIL import Image
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+from src.common.scroll_helper import ScrollHelper  # noqa: E402
+
+W, H = 64, 16
+
+
+def _items(rng, n):
+    out = []
+    for _ in range(n):
+        width = rng.randint(5, 60)
+        out.append(Image.frombytes("RGB", (width, H),
+                                   bytes(rng.randrange(256) for _ in range(width * H * 3))))
+    return out
+
+
+def _helper():
+    helper = ScrollHelper(W, H)
+    helper.create_scrolling_image(_items(random.Random(1), 4), item_gap=3, lead_gap=0)
+    return helper
+
+
+def _reference_append(strip, items, gap):
+    """What append_content used to do."""
+    width = sum(i.width for i in items) + gap * len(items)
+    addition = Image.new("RGB", (width, H))
+    x = 0
+    for item in items:
+        x += gap
+        addition.paste(item, (x, 0))
+        x += item.width
+    return np.concatenate((strip, np.array(addition)), axis=1)
+
+
+def test_an_append_writes_into_the_buffer_and_copies_only_the_new_columns():
+    helper = _helper()
+    helper.append_content(_items(random.Random(2), 2), item_gap=3)    # allocates
+    buffer = helper._strip_buffer
+    before = helper.cached_array.shape[1]
+    items = _items(random.Random(3), 2)
+    helper.append_content(items, item_gap=3)
+    assert helper._strip_buffer is buffer
+    assert np.shares_memory(helper.cached_array, buffer)
+    added = helper.cached_array.shape[1] - before
+    assert helper.last_copy_bytes == added * H * 3
+
+
+def test_a_trim_copies_nothing():
+    helper = _helper()
+    helper.append_content(_items(random.Random(2), 3), item_gap=3)
+    helper.scroll_position = 120.0
+    cut = helper.drop_scrolled_prefix()
+    assert cut == 120 and helper.last_copy_bytes == 0
+    assert np.shares_memory(helper.cached_array, helper._strip_buffer)
+
+
+def test_the_buffer_is_reallocated_when_the_room_runs_out():
+    helper = _helper()
+    helper.append_content(_items(random.Random(2), 1), item_gap=3)
+    first = helper._strip_buffer
+    rng = random.Random(4)
+    while helper._strip_buffer is first:
+        helper.append_content(_items(rng, 3), item_gap=3)
+    assert helper.last_copy_bytes == helper.cached_array.nbytes
+    assert helper._strip_start == 0
+
+
+def test_a_strip_set_from_outside_is_never_written_through():
+    # The multi-display follower adopts a read-only array straight from an image.
+    helper = _helper()
+    helper.append_content(_items(random.Random(2), 1), item_gap=3)
+    adopted = np.asarray(Image.new("RGB", (300, H), (9, 9, 9)))
+    helper.cached_array = adopted
+    helper.total_scroll_width = 300
+    helper.append_content(_items(random.Random(3), 1), item_gap=3)
+    assert not np.shares_memory(helper.cached_array, adopted)
+    assert (adopted == 9).all()
+    helper.cached_array = adopted
+    helper.scroll_position = 100.0
+    helper.drop_scrolled_prefix()
+    assert not np.shares_memory(helper.cached_array, adopted)
+
+
+def test_a_new_strip_lets_the_old_buffer_go():
+    helper = _helper()
+    helper.append_content(_items(random.Random(2), 1), item_gap=3)
+    helper.create_scrolling_image(_items(random.Random(5), 2), item_gap=3, lead_gap=0)
+    assert helper._strip_buffer is None
+    helper.append_content(_items(random.Random(2), 1), item_gap=3)
+    helper.clear_cache()
+    assert helper._strip_buffer is None and helper._strip_view is None
+
+
+@pytest.mark.parametrize("seed", range(12))
+def test_every_frame_matches_the_old_copying_strip(seed):
+    """Random appends, trims, scrolling and patches, against a strip kept the old way."""
+    rng = random.Random(seed)
+    helper = _helper()
+    reference = helper.cached_array.copy()
+    for _ in range(60):
+        op = rng.random()
+        if op < 0.35:
+            items = _items(rng, rng.randint(1, 3))
+            gap = rng.randint(0, 6)
+            helper.append_content(items, item_gap=gap)
+            reference = _reference_append(reference, items, gap)
+        elif op < 0.55:
+            keep = rng.randint(0, W)
+            before = helper.scroll_position
+            cut = helper.drop_scrolled_prefix(keep_before=keep)
+            reference = reference[:, cut:].copy()
+            assert helper.scroll_position == before - cut
+        elif op < 0.7 and helper.cached_array.shape[1] > 8:
+            x = rng.randrange(helper.cached_array.shape[1] - 4)
+            pixels = np.full((H, 4, 3), rng.randrange(256), dtype=np.uint8)
+            helper.patch_columns(x, pixels)
+            reference[:, x:x + 4] = pixels
+        else:
+            limit = max(0, helper.cached_array.shape[1] - W - 1)
+            helper.scroll_position = float(rng.randint(0, limit)) if limit else 0.0
+        assert helper.cached_array.shape == reference.shape
+        assert (helper.cached_array == reference).all()
+        assert helper.total_scroll_width == reference.shape[1]
+        frame = np.asarray(helper.get_visible_portion())
+        x = int(helper.scroll_position)
+        if x + W <= reference.shape[1]:
+            assert (frame == reference[:, x:x + W]).all()
+    # The lazily built image is the strip as it stands.
+    assert (np.asarray(helper.cached_image) == reference).all()