par2-rs

PAR2 verification and repair in pure Rust. No C bindings, no external par2 binary.

[dependencies]
par2-rs = "0.3"

Usage

use par2_rs::{DiskFileAccess, Par2FileSet, scan_packets_from_path, verify_all};

let packets = scan_packets_from_path("release.par2".as_ref())?
    .into_iter()
    .map(|(packet, _offset)| packet)
    .collect();
let set = Par2FileSet::from_packets(packets)?;

let access = DiskFileAccess::new("/downloads/release".into(), &set);
let result = verify_all(&set, &access);
println!("{} blocks missing", result.total_missing_blocks);

Par2Repairer drives the full sequence: scan, verify, solve, repair, verify again.

Capabilities

• All PAR2 packet types: Main, File Description, IFSC, Recovery Slice, Creator.
• Packets from any number of .par2 files aggregate into one set.
• Slice-level verification from IFSC CRC32 + MD5 pairs, so damage is localised rather than condemning a whole file.
• 16 KB quick-check for cheap file identification; full-file MD5 for sets with no IFSC data.
• Placement-aware repair: renamed and moved files are matched by content.
• Malformed or truncated packets are skipped by scanning forward, rather than failing the set.

Verifying data that is not on disk

verify_all reads through the FileAccess trait. DiskFileAccess is the ordinary implementation, but supplying your own allows verification against bytes that are still arriving over a network, or that are assembled from a source with no file paths at all.

Performance

rarpar 0.3.0 release validation

These deterministic end-to-end runs use the synthetic rarpar-bench corpus, one warmup, seven measured runs, canonical PAR2 placement, and SHA-256 output validation. Windows and Linux CLI builds are CPU-only. Apple Silicon is shown with CPU-only and Metal-capable builds. Normal runtime gating, not a force override, chooses Metal only for qualifying heavy repairs, while verification and smaller repairs remain on CPU. These runs include CLI discovery, repair, and post-repair verification rather than measuring only the repair kernel.

Broader library workloads

Measured against par2cmdline-turbo 1.4.0 on the same damaged sets, across an 11-scenario differential suite. Warm-cache medians from release builds with shipped flags; every repair is byte-compared against pristine output. Test machines: Apple M5 Max (macOS), Intel Core Ultra 9 285H (Ubuntu 24.04), AMD Ryzen 5 3600 (Windows, Zen 2 — no GFNI).

Repair

Verification

GPU repair (Apple Silicon, metal feature)

Verification wins on every machine: candidate windows are CRC-gated, MD5 is confirmed only when needed, and a single file is scanned in parallel where turbo's damaged scan is serial.

Repair is shape-dependent. The many-file case wins; the single-file, many-slice case still trails on x86 and Windows, where turbo's ParPar auto-selects an XOR-JIT kernel that runs the GF(2¹⁶) multiply on vector-XOR ports rather than the saturated shuffle ports. An equivalent tier is in progress.

Provenance

This is an independent Rust implementation, heavily informed by par2cmdline-turbo, Anime Tosho's speed-focused fork of par2cmdline. Both are GPL-2.0-or-later. That work is also the benchmark reference this crate is measured against.

The PAR2 format itself is specified in the Parity Volume Set Specification 2.0.

Versioned API and migration notes are in CHANGELOG.md.

License

GPL-3.0-or-later. See LICENSE.