## Summary
- **What is the goal of this PR?** Fix chapter-opener text collapsing to
1-2 words per line in EPUBs that apply large em-based horizontal CSS
insets.
- **What changes are included?** Caps `marginLeft`, `marginRight`,
`paddingLeft`, and `paddingRight` at 2em in `BlockStyle::fromCssStyle`
(`lib/Epub/Epub/blocks/BlockStyle.h`). Vertical margins/padding are
unchanged - the bug is horizontal-only.
## Additional Context
**Repro:** *Mother Night* by Kurt Vonnegut, Chapter 21 ("My best
friend..."). The chapter-opening element's embedded CSS sets a large
horizontal inset.
- Settings: Embedded Style = On, Justify alignment (default).
- Before: 1-2 words per line with a visible river between them.
- After: body text fills the usable page width like every other
paragraph.
**Why the clamp lives in `fromCssStyle`:** This is the single point
where CSS lengths resolve to pixels, so clamping here keeps both
`effectiveWidth` call sites in `ChapterHtmlSlimParser` (`:1131-1133`
makePages, `:841-844` long-block split) consistent with the
`leftInset()` xOffset. A width-only clamp at either call site would
leave text pushed right by a large xOffset.
**Test plan:**
- [x] Flashed to Xteink X4. Chapter 21 of *Mother Night* renders
normally with Embedded Style on.
- Users with cached layouts of affected books need to delete
`.crosspoint/epub_<hash>/sections/` (or the whole `.crosspoint/`) on the
SD card to pick up the fix.
**Before / After:**
<p float="left">
<img height="400" alt="IMG_0320"
src="https://github.com/user-attachments/assets/de8815ae-f2f4-4845-be1f-449c5a25d191"
/>
<img height="400" alt="IMG_0324"
src="https://github.com/user-attachments/assets/95f5f4cb-8aa1-418e-b611-18e0c342cbef"
/>
</p>
---
### 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 >**_
I used Claude Code (Opus 4.7) to evaluate the codebase and find the
relevant code related to this rendering. From there, it was human
designed, reviewed and tested.
Co-authored-by: rhoopr <>
## Summary
* **What is the goal of this PR?**
* Increase the allowed link length for footnotes
* **What changes are included?**
* Un-magic-numbers parts of the footnote parser
* Increases max href length
## Additional Context
* Additional RAM usage was not noticable to me, worst case 16 * 128 * 2
= 4KB instead of previously 16 * 88 * 2 ~ 2.8KB, see #1031
* Motivation: My book navigated me to the start of the footnote chapter,
which required manual navigating for every footnote due to ungodly file
name choices which pushed the required href length to 72, which caused
truncating, which caused Crosspoint to not find the anchor (Sacred and
Terrible Air, available via reddit)
---
### 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? _**NO**_
Switch KOReader sync progress mapping from chapter matching to
XPath-based mapping.
- resolves KOReader positions using real XHTML ancestry paths
- supports paragraph-based upload mapping with text offsets
where needed
- passes the current paragraph index into sync so uploads map
back to KOReader more accurately
No HTTP client changes are included. No reader-state or resume-flow
changes are included.
---------
Co-authored-by: jpirnay <jens@pirnay.com>
## Summary
**What is the goal of this PR?**
Added `destroyXmlParser()` helper to replace the repeated 4-line parser
cleanup block (stop, clear callbacks, free, null) that was copyied
across 6 XML parser files.
---
### 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? _**PARTIALLY**_
## Summary
Previouls where ALL whitespace (and square brackes) removed from the
footnote link text, however some link texts are multiworded, like "`turn
to 252`" which were truncated into "`turnto252`", (an example from the
first book "Flight from the Dark" of the Lone Wolf book series by Joe
Dever, see [link](https://www.projectaon.org/en/Main/FlightFromTheDark))
* This change will only remove whitespaces from the beginning and end of
the string
so "` [ 12 ] `" will become "`12`" just like before, and "` turn to 252
`" will become "`turn to 252`".
---
### 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? _**NO**_
## Summary
**What is the goal of this PR?** (e.g., Implements the new feature for
file uploading.)
`DESTRUCTOR_CLOSES_FILE=1` is set in platformio.ini, which makes SdFat's
FsBaseFile destructor call close() automatically when a file goes out of
scope.
Three categories of file close calls remain untouched:
1. Close before Storage.remove() on the same path: ScreenshotUtil.cpp
closes the file before deleting it on write error. The remove might fail
if the file is still open.
2. Close before reopening the same variable: Epub.cpp writes a temp
NCX/nav file, closes it, then reopens it for reading. The
RecentBooksStore.cpp close before saveToFile() is the same pattern, it
rewrites the same file.
3. Close on member variables: BookMetadataCache.cpp (bookFile,
spineFile, tocFile), Section.cpp (file), XtcParser.cpp (m_file),
ZipFile.cpp (file). These persist beyond any single function scope, so
the destructor timing doesn't match the intended close point.
---
### 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? _**PARTIALLY**_
## Summary
* **What is the goal of this PR?** Slightly increase the OOM limit when
loading epubs
* **What changes are included?** Switched vectors for parsing epubs to
deques, allowing use of more memory
## Additional Context
* Increases loadable epub size from 2000+ chapter/ToC entries loadable
to 5000+ chapter/ToC entries
* #1574, but without the complicated stuff
---
### 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? _**NO**_
### Testing
| Commit | Book | Time |
|----------|------|-------|
| 83cd96bc2f | 1000.epub | ~30 sec |
| 83cd96bc2f | 5000.epub | crash |
| PR | 1000.epub | ~29 sec |
| PR | 5000.epub | ~2 min 20 sec |
=> No actual loading time regressions
[tested_epubs.zip](https://github.com/user-attachments/files/26645243/tested_epubs.zip)
## Summary
Improves text antialiasing quality on the Xteink X3 (SSD1677) display to
bring it closer to X4 rendering quality. Addresses white lines through
letter strokes and ghosting artifacts during page turns and screen
transitions.
## Changes
### Display Driver (open-x4-sdk)
- Dedicated X3 grayscale LUTs with tuned VDL drive strengths for dark
gray (2 time units) and light gray (3 time units), with active GND hold
on non-gray transitions to prevent floating source crosstalk
- Tight scan timing: TP2/TP3 reduced to 1 (total gate-on 7 units vs 17),
minimizing parasitic charge leakage that caused white lines through
letter strokes
- Fast diff BB reinforcement: Added mild VDH reinforcing pulse to
lut_x3_bb_full so black pixels are actively driven during differential
refreshes, clearing gray residue/ghosting
- displayGrayBuffer() updated to use the dedicated gray LUT bank instead
of full refresh LUTs for X3
### Rendering Pipeline
- Re-enabled light gray rendering for X3 text and images, now safe with
dedicated gray LUTs providing proper 4-level gray
- Removed isLightGrayRestricted() gating that was limiting X3 to 3-level
gray
- Runtime display dimensions in DirectPixelWriter and ScreenshotUtil,
replaced hardcoded constants with runtime getters to support X3 792x528
resolution
### Note
The open-x4-sdk submodule references a commit on
juicecultus/community-sdk. A corresponding PR to
open-x4-epaper/community-sdk should be merged first so the submodule ref
resolves on upstream.
## Testing
Tested on physical X3 hardware. White lines through letters
significantly reduced, in-book ghosting improved via BB reinforcement,
antialiasing visually closer to X4 quality.
## AI Disclosure
Yes, AI was used to assist with the development of these changes.
---------
Replace hardcoded DISPLAY_WIDTH, DISPLAY_HEIGHT, and DISPLAY_WIDTH_BYTES
constants with runtime values from display object to support multiple
device models (X3 and X4) with different screen dimensions.
## Summary
**What is the goal of this PR?**
Replace verbose constructor initializer lists with in-class default
member initializers in JpegContext and PngContext
---
### 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? _**NO**_
## Summary
EPUBs that use ISO 639-2 three-letter language codes in their
`dc:language` metadata (e.g. `<dc:language>eng</dc:language>`) got no
hyphenation. The hyphenator registry only matched ISO 639-1 two-letter
codes (`"en"`, `"fr"`, etc.), so `"eng"` produced a null hyphenator and
every word in the book was treated as unhyphenatable.
Added a normalization step in `hyphenatorForLanguage` that maps ISO
639-2 codes (both bibliographic and terminological variants) to their
two-letter equivalents before the registry lookup.
Discovered via *Project Hail Mary* (Random House), which uses
`<dc:language>eng</dc:language>`.
---
### 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? _**PARTIALLY**_
## Summary
Adds Xteink X3 hardware support to CrossPoint Reader. The X3 uses the
same SSD1677 e-ink controller as the X4 but with a different panel
(792x528 vs 800x480), different button layout, and an I2C fuel gauge
(BQ27220) instead of ADC-based battery reading.
All X3-specific behavior is gated by runtime device detection — X4
behavior is unchanged.
Depends on community-sdk X3 support: open-x4-epaper/community-sdk#19
(merged).
## Changes
### HAL Layer
**HalGPIO** (`lib/hal/HalGPIO.cpp/.h`)
- I2C-based device fingerprinting at boot: probes for BQ27220 fuel
gauge, DS3231 RTC, and QMI8658 IMU to distinguish X3 from X4
- Detection result cached in NVS for fast subsequent boots
- Exposes `deviceIsX3()` / `deviceIsX4()` helpers used throughout the
codebase
- X3 button mapping (7 GPIOs vs X4's layout)
- USB connection detection and wake classification for X3
**HalDisplay** (`lib/hal/HalDisplay.cpp/.h`)
- Calls `einkDisplay.setDisplayX3()` before init when X3 is detected
- Requests display resync after power button / flash wake events
- Runtime display dimension accessors (`getDisplayWidth()`,
`getDisplayHeight()`, `getBufferSize()`)
- Exposed as global `display` instance for use by image converters
**HalPowerManager** (`lib/hal/HalPowerManager.cpp/.h`)
- X3 battery reading via I2C fuel gauge (BQ27220 at 0x55, SOC register)
- X3 power button uses GPIO hold for deep sleep
### Display & Rendering
**GfxRenderer** (`lib/GfxRenderer/GfxRenderer.cpp/.h`)
- Buffer size and display dimensions are now runtime values (not
compile-time constants) to support both panel sizes
- X3 anti-aliasing tuning: only the darker grayscale level is applied to
avoid washed-out text on the X3 panel. X4 retains both levels via
`deviceIsX4()` gate
**Image Converters** (`lib/JpegToBmpConverter`, `lib/PngToBmpConverter`)
- Cover image prescale target uses runtime display dimensions from HAL
instead of hardcoded 800x480
### UI Themes
**BaseTheme / LyraTheme** (`src/components/themes/`)
- X3 button position mapping for the different physical layout
- Adjusted UI element positioning for 792x528 viewport
### Boot & Init
**main.cpp**
- X3 hardware detection logging
- Adjusted init sequence for X3 (no `HalSystem::begin()` dependency on
X3 path)
**HomeActivity**
- Uses runtime `renderer.getBufferSize()` instead of static
`GfxRenderer::getBufferSize()`
FYI I did not add support for the gyro page turner. That can be it's own
PR.
## Summary
Replace per-pixel getRenderMode() + rotateCoordinates() + bounds checks
with a DirectPixelWriter struct that pre-computes orientation and render
mode state once per row. Use bitwise ops instead of division/modulo for
cache pixel packing. Skip PNG cache allocation when buffer exceeds 48KB
(framebuffer size) since PNG decode is fast enough that caching provides
minimal benefit, and the large buffer competes with the 44KB PNG decoder
for heap.
## Additional Context
Measured improvements on ESP32-C3 @ 160MHz:
- JPEG decode: 5-7% faster (1:1 scale)
- PNG decode: 15-20% faster (1:1 scale)
- Cache renders: 3-6% faster across both formats
- Eliminates "Failed to allocate cache buffer" errors for large PNGs
---
### 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? _**< PARTIALLY >**_
## Summary
- Add CSS `display: none` support to the EPUB rendering pipeline (fixes
#1431)
- Parse `display` property in stylesheets and inline styles, with full
cascade resolution (element, class, element.class, inline)
- Skip hidden elements and all their descendants in
`ChapterHtmlSlimParser`
- Separate display:none check for `<img>` tags (image code path is
independent of the general element handler)
- Flush pending text blocks before placing images to fix layout ordering
(text preceding an image now correctly renders above it)
- Bump CSS cache version to 4 to invalidate stale caches
- Add test EPUB (`test_display_none.epub`) covering class selectors,
element selectors, combined selectors, inline styles, nested hidden
content, hidden images, style priority/override, and realistic use cases
## Summary
* **What is the goal of this PR?** Fixing / extending the hyphenation
logic to deal with words containing an apostophe as raised in #1186
* **What changes are included?**
## Additional Context
---
### 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? _**PARTIALLY**_ (as the
user provided a thorough analysis that I followed)
## Purpose
This PR includes some preparatory changes that are needed for an
upcoming performant CJK font feature. The changes have no impact on
render time and heap allocation for latin text. **Despite this, I think
these changes stand on their own as a better font
compression/decompression implementation.**
## Summary
- Font decompressor rewrite: Replaced the 4-slot LRU group cache with a
two-tier system — a page buffer (glyphs prewarmed before rendering
begins) and a hot-group fallback (last decompressed group retained for
non-prewarmed
glyphs).
- Byte-aligned compressed bitmap format: Glyph bitmaps within compressed
groups are now stored row-padded rather than tightly packed before
DEFLATE compression, improving compression ratios by making identical
pixel rows produce
identical byte patterns. Glyphs are compacted back to packed format on
demand at render time. Reduces flash size by 155 KB.
- Page prewarm system: Added `Page::collectText` and
`Page::getDominantStyle` to extract per-style glyph requirements before
rendering, and `GfxRenderer::prewarmFontCache` to pre-decompress only
the groups needed for the dominant style
— eliminating mid-render decompression for the common case.
- UTF-8 robustness fixes: `utf8NextCodepoint` now validates continuation
bytes and returns a replacement glyph on malformed input;
`ChapterHtmlSlimParser` correctly preserves incomplete multi-byte
sequences across word-buffer flush
boundaries rather than splitting them.
---
### 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**_ Architecture and
design was done by me, refined a bit by Claude. Code mostly by Claude,
but not entirely.
## Summary
**What is the goal of this PR?**
### Problem
Inter-word gap widths were computed as two separately-snapped integers:
```cpp
gap = getSpaceWidth(fontId, style); // fp4::toPixel(spaceAdvance)
gap += getSpaceKernAdjust(fontId, leftCp, rightCp, style); // fp4::toPixel(kern1 + kern2)
```
Because `fp4::toPixel(a) + fp4::toPixel(b)` can differ from
`fp4::toPixel(a + b)` by +/-1 pixel when the fractional parts straddle a
rounding boundary, each inter-word space could be one pixel wider or
narrower than the correct value. This affected line-break width
decisions and word-position accumulation across the whole paragraph
layout pipeline.
### Fix
Replaces `getSpaceKernAdjust()` with `getSpaceAdvance(fontId, leftCp,
rightCp, style)`, which combines the space glyph advance and both
flanking kern values (`kern(leftCp, ' ')` + `kern(' ', rightCp)`) into a
single fixed-point sum before the snap:
```cpp
return fp4::toPixel(spaceAdvanceFP + kern(leftCp, ' ') + kern(' ', rightCp));
```
This is the same single-snap pattern already used by `getTextAdvanceX`
for word widths.
### Changes
- **`GfxRenderer`**: Replaces `getSpaceKernAdjust()` with
`getSpaceAdvance()`. `getSpaceWidth()` is retained for the
single-space-word case in `measureWordWidth` where no adjacent-word kern
context is available.
- **`ParsedText`**: All four call sites (`computeLineBreaks`,
`computeHyphenatedLineBreaks`, and both loops in `extractLine`) updated
to use `getSpaceAdvance()`. The now-redundant `spaceWidth`
pre-computation and parameter are removed from all three internal layout
functions.
---
### 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 to analyze for
correctness**_
## Summary
Fixes a crash (load access fault) when opening EPUB chapters whose first
text block exceeds 750 words.
## Changes
* **Crash fix (`addLineToPage`)**: Added a null guard for `currentPage`.
If `makePages()` hasn't run yet (which can happen when the first block
triggers the "text block too long" split path), the page is now created
on demand.
* **Layout fix (`characterData`)**: The early-split path previously used
`viewportWidth`, ignoring CSS margins and padding. It now computes
`effectiveWidth` using `totalHorizontalInset()`, consistent with
`makePages()`.
## Additional Context
* Crash signature: `MCAUSE=0x5` (load access fault), `A0=0x0` (`Page*`
null), `MTVAL=0x4 / 0x8` (offsets into `Page::elements`).
* Confirmed in two user reports reported in #1328
* Tested on PR #1357 (not on `master`).
---
### 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? _**< PARTIALLY >**_
## Summary: Enable footnote anchor navigation in EPUB reader
This PR extracts the core anchor-to-page mapping mechanism from PR #1143
(TOC fragment navigation) to provide immediate footnote navigation
support. By merging this focused subset first, users get a complete
footnote experience now while simplifying the eventual review and merge
of the full #1143 PR.
---
## What this extracts from PR #1143
PR #1143 implements comprehensive TOC fragment navigation for EPUBs with
multi-chapter spine files. This PR takes only the anchor resolution
infrastructure:
- Anchor-to-page mapping in section cache: During page layout,
ChapterHtmlSlimParser records which page each HTML id attribute lands
on, serializing the map into the .bin cache file.
- Anchor resolution in `EpubReaderActivity`: When navigating to a
footnote link with a fragment (e.g., `chapter2.xhtml#note1`), the reader
resolves the anchor to a page number and jumps directly to it.
- Section file format change: Bumped to version 15, adds anchor map
offset in header.
---
## Simplified scope vs. PR #1143
To minimize conflicts and complexity, this PR differs from #1143 in key
ways:
* **Anchors tracked**
* **Origin:** Only TOC anchors (passed via `std::set`)
* **This branch:** All `id` attributes
* **Page breaks**
* **Origin**: Forces new page at TOC chapter boundaries
* **This branch:** None — natural flow
* **TOC integration**
* **Origin**: `tocBoundaries`, `getTocIndexForPage()`, chapter skip
* **This branch:** None — just footnote links
* **Bug fix**
* **This branch:** Fixed anchor page off-by-1/2 bug
The anchor recording bug (recording page number before `makePages()`
flushes previous block) was identified and fixed during this extraction.
The fix uses a deferred `pendingAnchorId` pattern that records the
anchor after page completion.
---
## Positioning for future merge
Changes are structured to minimize conflicts when #1143 eventually
merges:
- `ChapterHtmlSlimParser.cpp` `startElement()`: Both branches rewrite
the same if `(!idAttr.empty())` block. The merged version will combine
both approaches (TOC anchors get page breaks + immediate recording;
footnote anchors get deferred recording).
- `EpubReaderActivity.cpp` `render()`: The `pendingAnchor` resolution
block is positioned at the exact same insertion point where #1143 places
its `pendingTocIndex` block (line 596, right after `nextPageNumber`
assignment). During merge, both blocks will sit side-by-side.
---
## Why merge separately?
1. Immediate user value: Footnote navigation works now without waiting
for the full TOC overhaul
2. Easier review: ~100 lines vs. 500+ lines in #1143
3. Bug fix included: The page recording bug is fixed here and will carry
into #1143
4. Minimal conflicts: Structured for clean merge — both PRs touch the
same files but in complementary ways
---
### AI Usage
Did you use AI tools to help write this code? _**< YES >**_ Done by
Claude Opus 4.6
## Summary
**What is the goal of this PR?**
This change avoids the pattern of creating a `std::string` using
`.substr` in order to compare against a file extension literal.
```c++
std::string path;
if (path.length() >= 4 && path.substr(path.length() - 4) == ".ext")
```
The `checkFileExtension` utility has moved from StringUtils to
FsHelpers, to be available to code in lib/. The signature now accepts a
`std::string_view` instead of `std::string`, which makes the single
implementation reusable for Arduino `String`.
Added utility functions for commonly repeated extensions.
These changes **save about 2 KB of flash (5,999,427 to 5,997,343)**.
---
### 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? _**NO**_
## Summary
**What is the goal of this PR?**
Avoid building cache path strings twice, once to check existence of the
file and a second time to delete the file.
---
### 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? _**NO**_
## Summary
**What is the goal of this PR?** Add a user setting to decide image
support: display, show placeholder instead, supress fully
Fixes#1289
---
### 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? _**< NO >**_
## Summary
* **What is the goal of this PR?** Fixing two independent CSS rendering
bugs combined to make hanging-indent list styles
(e.g. margin-left:3em; text-indent:-1em) render incorrectly:
* **What changes are included?**
1. Negative text-indent was silently ignored
Three guards in ParsedText.cpp (computeLineBreaks,
computeHyphenatedLineBreaks,
extractLine) conditioned firstLineIndent on blockStyle.textIndent > 0,
so any
negative value collapsed to zero. Additionally, wordXpos was uint16_t,
which
cannot represent negative offsets — a cast of e.g. −18 would wrap to
65518 and
render the word far off-screen.
2. extraParagraphSpacing suppressed hanging indents
Even after removing the > 0 guard, the existing !extraParagraphSpacing
condition
would still suppress all text-indent when that setting is on (its
default). Positive
text-indent is a decorative paragraph indent that the user can
reasonably replace with
vertical spacing — negative text-indent is structural (it positions the
list marker)
and must always apply.
3. em unit was calibrated against line height, not font size
emSize was computed as getLineHeight() * lineCompression (the full line
advance).
CSS em units are defined relative to the font-size, which corresponds to
the
ascender height — not the line height. Using line height makes every
em-based
margin/indent ~20–30% wider than a browser would render it, and is
especially
noticeable for CSS that uses font-size: small (which we do not
implement).
## Additional Context
Test case
```
.lsl1 { margin-left: 3em; text-indent: -1em; }
<div class="lsl1">• First list item that wraps across lines</div>
<div class="lsl1">• Short item</div>
```
Before: all lines of all items started at 3 em from the left edge
(indent ignored).
After: the bullet marker hangs at 2 em; continuation lines align at 3
em.
<img width="240" alt="before"
src="https://github.com/user-attachments/assets/9dcbf3e0-fcd9-4af8-b451-a90ba4d2fb75"
/>
<img width="240" alt="after"
src="https://github.com/user-attachments/assets/1ffdcf56-a180-4267-9590-c60d7ac44707"
/>
---
### 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**_
## Summary
Replaces the picojpeg library with bitbank2/JPEGDEC for JPEG decoding in
the EPUB image pipeline. JPEGDEC provides built-in coarse scaling (1/2,
1/4, 1/8), 8-bit grayscale output, and streaming block-based decoding
via callbacks.
Includes a pre-build patch script for two JPEGDEC changes affecting
progressive JPEG support and EIGHT_BIT_GRAYSCALE mode.
Closes#912
## Additional Context
# Example progressive jpeg
<img
src="https://github.com/user-attachments/assets/e63bb4f8-f862-4aa0-a01f-d1ef43a4b27a"
width="400" height="800" />
Good performance increase from JPEGDEC over picojpeg cc @bitbank2 thanks
## Baseline JPEG Decode Performance: picojpeg vs JPEGDEC (float in
callback) vs JPEGDEC (fixed-point in callback)
Tested with `test_jpeg_images.epub` on device (ESP32-C3), first decode
(no cache).
| Image | Source | Output | picojpeg | JPEGDEC float | JPEGDEC
fixed-point | vs picojpeg | vs float |
|-------|--------|--------|----------|---------------|---------------------|-------------|----------|
| jpeg_format.jpg | 350x250 | 350x250 | 313 ms | 256 ms | **104 ms** |
**3.0x** | **2.5x** |
| grayscale_test.jpg | 400x600 | 400x600 | 768 ms | 661 ms | **246 ms**
| **3.1x** | **2.7x** |
| gradient_test.jpg | 400x500 | 400x500 | 707 ms | 597 ms | **247 ms** |
**2.9x** | **2.4x** |
| centering_test.jpg | 350x400 | 350x400 | 502 ms | 412 ms | **169 ms**
| **3.0x** | **2.4x** |
| scaling_test.jpg | 1200x1500 | 464x580 | 5487 ms | 1114 ms | **668
ms** | **8.2x** | **1.7x** |
| wide_scaling_test.jpg | 1807x736 | 464x188 | 4237 ms | 642 ms | **497
ms** | **8.5x** | **1.3x** |
| cache_test_1.jpg | 400x300 | 400x300 | 422 ms | 348 ms | **141 ms** |
**3.0x** | **2.5x** |
| cache_test_2.jpg | 400x300 | 400x300 | 424 ms | 349 ms | **142 ms** |
**3.0x** | **2.5x** |
### Summary
- **1:1 scale (fixed-point vs float)**: ~2.5x faster — eliminating
software float on the FPU-less ESP32-C3 is the dominant win
- **1:1 scale (fixed-point vs picojpeg)**: ~3.0x faster overall
- **Downscaled images (vs picojpeg)**: 8-9x faster — JPEGDEC's coarse
scaling + fixed-point draw callback
- **Downscaled images (fixed-point vs float)**: 1.3-1.7x — less dramatic
since JPEG library decode dominates over the draw callback for fewer
output pixels
- The fixed-point optimization alone (vs float JPEGDEC) saved **~60% of
render time** on 1:1 images, confirming that software float emulation
was the primary bottleneck in the draw callback
- See thread for discussions on quality of progressive images,
https://github.com/crosspoint-reader/crosspoint-reader/pull/1136#issuecomment-3952952315
- and the conclusion
https://github.com/crosspoint-reader/crosspoint-reader/pull/1136#issuecomment-3959379386
- Proposal to improve quality added at
https://github.com/crosspoint-reader/crosspoint-reader/discussions/1179
---
### 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? _**< PARTIALLY >**_
---------
Co-authored-by: Dave Allie <dave@daveallie.com>
CSCMeandGitHub
Zach NelsonandGitHub
Stefan Blixten KarlssonandGitHub
JustinianandGitHub
martin brookandGitHub
Adrian Wilkins-CaruanaandGitHub
Uri TauberandGitHub
