Introduces non-blocking display refresh methods that allow CPU work (like grayscale rendering) to overlap with the e-ink panel's refresh time. The async path starts the waveform and returns immediately, with the caller responsible for waiting via waitRefreshComplete(). Falls back to blocking refresh when fadingFix is enabled or the panel lacks deferral support.
Temporarily release framebuffer memory during section pagination to reduce heap pressure. The framebuffer can be released before memory-intensive operations and restored afterward, allowing BLE stack startup threshold to be lowered from 80KB to 70KB. Implements RAII-style FrameBufferBuildLoan class to ensure proper cleanup and automatic restart on restoration failure.
Replace per-page buffer free+realloc pattern with capacity tracking that only reallocates when needed size exceeds current capacity. This prevents heap fragmentation from non-coalescing holes that occurred when each page's freed block rarely fit the next page's allocation. After a few page turns, capacities converge on the book's maximum and page turns stop touching the allocator entirely.
Prevent OOM aborts by shedding BLE stack when free heap drops below 24KB before rendering. Lower BLE start threshold from 100KB to 80KB to match steady-state heap levels (~84KB). Reserve vector capacity upfront in page parsing to avoid reallocation crashes on fragmented heaps. Gate background section builds on heap availability.
Field session: after a recovery teardown the heap settles at ~72 KB --
below the 100 KB start floor -- and never recovers on its own (the
defrag silent-restart only fires when a build FAILS, and builds succeed
now). BT stayed paused forever, the menu toggle claimed ON, and toggling
off/on changed nothing.
- Menu label tells the truth: ON / PAUSED (enabled but stack down) / OFF
(new STR_STATE_PAUSED string)
- Toggling BT on below the heap floor silent-restarts into the current
book: the fresh boot's ~118 KB passes the gate and BT auto-starts on
resume. Explicit user intent is the right trigger for the defrag.
- Hoist the floor to bleinput::kStartMinFreeHeap, shared by the
lifecycle gate and the toggle
## Summary
* **What is the goal of this PR?** Fixes#2402. On the X3, the "Time to
Sleep"
picker's 5-minute side buttons were inverted (left increased, right
decreased)
and the on-screen legend didn't match the physical buttons.
* **What changes are included?**
* Flip the large-step (±5 min) direction on X3 so the left side button
decreases and the right increases, matching the layout. X4 is unchanged.
* Add a device-specific step-hint string
(`STR_SLEEP_TIMER_STEP_HINT_X3`)
selected via `gpio.deviceIsX3()`. X3 shows `+/-: 1 min Side buttons: 5
min`;
X4 keeps the original `Left/Right: 1 min Up/Down: 5 min`.
* Same fix applied to both slider pickers that read the raw up/down side
buttons: the Time-to-Sleep picker and the "Go to %" picker
(`EpubReaderPercentSelectionActivity`), each with its own X3 hint string
(`STR_SLEEP_TIMER_STEP_HINT_X3`, `STR_PERCENT_STEP_HINT_X3`).
## Additional Context
* **Root cause:** the X3's side buttons sit one on each edge of the
screen
(power on top), whereas the X4 has a vertical up/down rocker on the
right
edge. So `BTN_UP` is physically the *left* button on X3 but the *top*
button
on X4. The picker mapped up→+5 / down→−5 unconditionally, which reads
naturally on the X4 rocker but inverted on the X3's left/right buttons.
The
static legend ("Up/Down: 5 min") had the same X3-only mismatch.
* The codebase already detects the device at runtime and handles this
layout
difference elsewhere (e.g. `LyraTheme::drawSideButtonHints` draws "Up on
left,
Down on right" for X3), so this reuses the same `gpio.deviceIsX3()`
signal.
* Swept the codebase for other side-button sliders:
`ClockOffsetActivity` already
uses the `Next/Previous` (`ButtonNavigator`) abstraction, which resolves
to
left=decrement / right=increment on X3, so it needs no change. List/page
navigation (Wi-Fi, KOReader sync, BMP viewer) and the keyboard cursor
are also
unaffected. The two slider pickers above were the only ones using raw
up/down.
* **X4 is untouched** — same direction, same legend wording as before.
* **Tested on X3 hardware:** left side button now decreases, right
increases, and
the legend matches. X4 not yet tested on device (no unit on hand); its
behaviour and strings are unchanged from master.
* **Translations:** the new `STR_SLEEP_TIMER_STEP_HINT_X3` and
`STR_PERCENT_STEP_HINT_X3` were added to all 26 languages, but the
non-English
ones are AI-generated and would benefit from a native-speaker pass —
particularly Hebrew (RTL ordering with a leading `+/-`), Kazakh, and
Vietnamese.
---
### AI Usage
Did you use AI tools to help write this code? _**YES**_
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
## Summary
* **What is the goal of this PR?**
Replace the pixel-by-pixel `fillRectDither` implementation with a new
byte-aligned `fillRectImpl` that eliminates per-pixel
`rotateCoordinates` calls and Read-Modify-Write bitwise loops, yielding
a significant rendering speedup on ESP32 E-ink framebuffers.
Ported from @rhythmerc's crosspoint-reader fork commit 27ea625.
* **What changes are included?**
- **`GfxRenderer.cpp` — `fillRectDither` refactor:** The existing
`if/else if` chain is replaced with a `switch` statement that delegates
each `Color` case to the new `fillRectImpl<Color>()` template,
eliminating runtime branching.
- **`GfxRenderer.cpp` — new `fillRectImpl<C>()` template:** Core of the
optimization. Key behaviors:
- Clips the rectangle in logical space upfront.
- Rotates only **2 opposing corner points** (top-left and bottom-right)
into physical framebuffer space instead of rotating every pixel
individually.
- Derives physical-space `byteStart`/`byteEnd` and precomputes
`headMask` / `tailMask` for MSB-first partial-byte boundaries,
performing RMW only on the edge bytes.
- **Solid fills (`Black` / `White`):** Uses `memset` for all interior
full-byte runs per row — no per-pixel writes.
- **Dithered fills (`LightGray` / `DarkGray`):** Precomputes both parity
variants of `blackMask` (even/odd `py`) **outside** the row loop,
eliminating the previously re-evaluated 8-bit construction loop on every
physical row. Interior full bytes are then written with a single
`memset(whiteMask)`.
- Uses `if constexpr` throughout to dispatch on `Color` at compile time,
generating zero runtime branches per template instantiation.
- **`GfxRenderer.h`:** Declares the new private `fillRectImpl<Color>()`
template method with an explanatory doc-comment.
- **Explicit template instantiations** added for all four active `Color`
variants (`Black`, `White`, `LightGray`, `DarkGray`).
## Additional Context
* **Performance:** The primary motivation is ESP32 E-ink framebuffer
performance. The old path called `rotateCoordinates` and did a full RMW
for every single pixel in the rectangle. The new path calls
`rotateCoordinates` exactly **twice** per fill regardless of rectangle
size, then operates at byte granularity — a complexity reduction from
O(W×H) coordinate transforms to O(1).
* **Dither correctness:** The `blackMask` precomputation relies on the
dither pattern having period 2 in both logical X and Y, which makes the
per-row byte pattern repeat with period 2 in `py`. Reviewers should
verify the `lxBase`/`lyBase` derivations for all four orientations
(`Portrait`, `PortraitInverted`, `LandscapeClockwise`,
`LandscapeCounterClockwise`) match the inverse of `rotateCoordinates`.
* **Edge case — single-byte rows:** When `byteStart == byteEnd`, the
head and tail masks are ANDed together into a single `rectMask` to avoid
double-masking the same byte. This path should be tested with narrow
rectangles (width < 8px).
* **No behavioral change for `Color::Clear`:** The `Clear` case exits
early via `if constexpr` and is a no-op, matching the original behavior.
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**< YES >**_
---------
Co-authored-by: Ryan Mercado <rmercado@firstdollar.com>
Justin Mitchell