Some EPUB converters emit byte-identical stylesheets per chapter (100+ entries). Now scan the ZIP central directory once to identify duplicates by CRC32 and compressed size, then skip parsing identical files. This avoids redundant ZIP lookups and SD extraction round-trips while preserving all styles since rules merge into a global set. Also add extractItemToFile helper and allowEarlyStop parameter to readItemContentsToStream.
## Summary
* **What is the goal of this PR?**
* This PR ports over Crossink's handling of large EPUBs that helps
prevent crashes during open after the book metadata cache is built.
* **What changes are included?**
* Removes the post-indexing ZIP-wide CSS discovery pass that built an
in-memory map of every ZIP entry.
* Reuses `content.opf` parsing to collect declared CSS files without
writing spine entries again.
* Temporarily releases the loaded book metadata cache while rebuilding
CSS for cached books.
* Parses CSS before reloading `book.bin` after a fresh cache build,
leaving more heap available during CSS rule parsing.
## Additional Context
* User reported their EPUB opening fine on Crossink but would crash on
Crosspoint. Verified this claim on my own devices.
* The crash this addresses happened after `book.bin` was successfully
built, when CSS discovery allocated a large `unordered_map` for ~3k EPUB
ZIP entries.
* Tradeoff: CSS files not declared in `content.opf` are no longer
discovered by scanning the full ZIP. This avoids the high-risk memory
allocation but improperly formatted EPUBs (ones that don't declare their
CSS styles in `content.opf` will render without styling and fallback to
inline styles.
* User provided epub that was crashing prior to this change:
https://www.mediafire.com/file/g57ea4mj13iunvh/Quang+%C3%82m+Chi+Ngo%E1%BA%A1i+-+Nh%C4%A9+C%C4%83n.epub/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? _**< YES >**_
## 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
Continue my changes to introduce the HAL infrastructure from
https://github.com/crosspoint-reader/crosspoint-reader/pull/522
This PR touches quite a lot of files, but most of them are just name
changing. It should not have any impacts to the end behavior.
## Additional Context
My plan is to firstly add this small shim layer, which sounds useless at
first, but then I'll implement an emulated driver which can be helpful
for testing and for development.
Currently, on my fork, I'm using a FS driver that allow "mounting" a
local directory from my computer to the device, much like the `-v` mount
option on docker. This allows me to quickly reset `.crosspoint`
directory if anything goes wrong. I plan to upstream this feature when
this PR get merged.
---
### 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
Optimizes EPUB metadata indexing for large books (2000+ chapters) from
~30 minutes to ~50 seconds by replacing O(n²) algorithms with O(n log n)
hash-indexed lookups.
Fixes#134
## Problem
Three phases had O(n²) complexity due to nested loops:
| Phase | Operation | Before (2768 chapters) |
|-------|-----------|------------------------|
| OPF Pass | For each spine ref, scan all manifest items | ~25 min |
| TOC Pass | For each TOC entry, scan all spine items | ~5 min |
| buildBookBin | For each spine item, scan ZIP central directory | ~8.4
min |
Total: **~30+ minutes** for first-time indexing of large EPUBs.
## Solution
Replace linear scans with sorted hash indexes + binary search:
- **OPF Pass**: Build `{hash(id), len, offset}` index from manifest,
binary search for each spine ref
- **TOC Pass**: Build `{hash(href), len, spineIndex}` index from spine,
binary search for each TOC entry
- **buildBookBin**: New `ZipFile::fillUncompressedSizes()` API - single
ZIP central directory scan with batch hash matching
All indexes use FNV-1a hashing with length as secondary key to minimize
collisions. Indexes are freed immediately after each phase.
## Results
**Shadow Slave EPUB (2768 chapters):**
| Phase | Before | After | Speedup |
|-------|--------|-------|---------|
| OPF pass | ~25 min | 10.8 sec | ~140x |
| TOC pass | ~5 min | 4.7 sec | ~60x |
| buildBookBin | 506 sec | 34.6 sec | ~15x |
| **Total** | **~30+ min** | **~50 sec** | **~36x** |
**Normal EPUB (87 chapters):** 1.7 sec - no regression.
## Memory
Peak temporary memory during indexing:
- OPF index: ~33KB (2770 items × 12 bytes)
- TOC index: ~33KB (2768 items × 12 bytes)
- ZIP batch: ~44KB (targets + sizes arrays)
All indexes cleared immediately after each phase. No OOM risk on
ESP32-C3.
## Note on Threshold
All optimizations are gated by `LARGE_SPINE_THRESHOLD = 400` to preserve
existing behavior for small books. However, the algorithms work
correctly for any book size and are faster even for small books:
| Book Size | Old O(n²) | New O(n log n) | Improvement |
|-----------|-----------|----------------|-------------|
| 10 ch | 100 ops | 50 ops | 2x |
| 100 ch | 10K ops | 800 ops | 12x |
| 400 ch | 160K ops | 4K ops | 40x |
If preferred, the threshold could be removed to use the optimized path
universally.
## Testing
- [x] Shadow Slave (2768 chapters): 50s first-time indexing, loads and
navigates correctly
- [x] Normal book (87 chapters): 1.7s indexing, no regression
- [x] Build passes
- [x] clang-format passes
## Files Changed
- `lib/Epub/Epub/parsers/ContentOpfParser.h/.cpp` - OPF manifest index
- `lib/Epub/Epub/BookMetadataCache.h/.cpp` - TOC index + batch size
lookup
- `lib/ZipFile/ZipFile.h/.cpp` - New `fillUncompressedSizes()` API
- `lib/Epub/Epub.cpp` - Timing logs
<details>
<summary><b>Algorithm Details</b> (click to expand)</summary>
### Phase 1: OPF Pass - Manifest to Spine Lookup
**Problem**: Each `<itemref idref="ch001">` in spine must find matching
`<item id="ch001" href="...">` in manifest.
```
OLD: For each of 2768 spine refs, scan all 2770 manifest items
= 7.6M string comparisons
NEW: While parsing manifest, build index:
{ hash("ch001"), len=5, file_offset=120 }
Sort index, then binary search for each spine ref:
2768 × log₂(2770) ≈ 2768 × 11 = 30K comparisons
```
### Phase 2: TOC Pass - TOC Entry to Spine Index Lookup
**Problem**: Each TOC entry with `href="chapter0001.xhtml"` must find
its spine index.
```
OLD: For each of 2768 TOC entries, scan all 2768 spine entries
= 7.6M string comparisons
NEW: At beginTocPass(), read spine once and build index:
{ hash("OEBPS/chapter0001.xhtml"), len=25, spineIndex=0 }
Sort index, binary search for each TOC entry:
2768 × log₂(2768) ≈ 30K comparisons
Clear index at endTocPass() to free memory.
```
### Phase 3: buildBookBin - ZIP Size Lookup
**Problem**: Need uncompressed file size for each spine item (for
reading progress). Sizes are in ZIP central directory.
```
OLD: For each of 2768 spine items, scan ZIP central directory (2773 entries)
= 7.6M filename reads + string comparisons
Time: 506 seconds
NEW:
Step 1: Build targets from spine
{ hash("OEBPS/chapter0001.xhtml"), len=25, index=0 }
Sort by (hash, len)
Step 2: Single pass through ZIP central directory
For each entry:
- Compute hash ON THE FLY (no string allocation)
- Binary search targets
- If match: sizes[target.index] = uncompressedSize
Step 3: Use sizes array directly (O(1) per spine item)
Total: 2773 entries × log₂(2768) ≈ 33K comparisons
Time: 35 seconds
```
### Why Hash + Length?
Using 64-bit FNV-1a hash + string length as a composite key:
- Collision probability: ~1 in 2⁶⁴ × typical_path_lengths
- No string storage needed in index (just 12-16 bytes per entry)
- Integer comparisons are faster than string comparisons
- Verification on match handles the rare collision case
</details>
---
_AI-assisted development. All changes tested on hardware._
## Summary
* Swap to updated SDCardManager which uses SdFat
* Add exFAT support
* Swap to using FsFile everywhere
* Use newly exposed `SdMan` macro to get to static instance of
SDCardManager
* Move a bunch of FsHelpers up to SDCardManager
Still a bit raw, but gets the time required to determine the size of
each chapter (for reading progress) down from ~25ms to 0-1ms.
This is done by keeping the zipArchive open (so simple ;)).
Probably we don't need to cache the spine sizes anymore then...
---------
Co-authored-by: Dave Allie <dave@daveallie.com>
Justin Mitchell