fcpw

fcpw is a variation-tolerant CBOR codec for Rust that aims to be the fastest available CBOR parser in the general case. The name is an acronym for the Fastest CBOR Parser in the West. The name and phrase are an allusion to FFTW, the Fastest Fourier Transform in the West, because I'm not particularly original. Performance is not pursued at the expense of reliability; correctness and fault tolerance are equal design goals.

FCPW provides a zero-copy slice decoder, structural event parser, dynamic borrowed and owned values, Serde integration, deterministic encoding and validation, synchronous I/O, RFC 8742 sequence iteration, diagnostic notation, and opt-in parallel batch decoding.

use serde::{Deserialize, Serialize};

#[derive(Debug, PartialEq, Serialize, Deserialize)]
struct Message<'a> {
    id: u64,
    text: &'a str,
}

let message = Message { id: 7, text: "hello" };
let bytes = fcpw::to_vec(&message)?;
let decoded: Message<'_> = fcpw::from_slice(&bytes)?;
assert_eq!(decoded, message);
# Ok::<(), fcpw::Error>(())

Normal encoding preserves the source float width: f32 is emitted as CBOR binary32 and f64 as binary64. Deterministic encoding uses the shortest exact float width and canonicalizes NaNs. The low-level Encoder exposes the same choice through f32/f64 and f32_preferred/f64_preferred.

Repeated encoding can retain output capacity with to_vec_into or to_vec_deterministic_into:

let mut output = Vec::with_capacity(1024);
fcpw::to_vec_into(&message, &mut output)?;
// The next call clears the bytes but reuses the allocation.
fcpw::to_vec_into(&message, &mut output)?;
# Ok::<(), fcpw::Error>(())

Repeated reader decoding can likewise retain its input allocation with from_reader_with_buffer. Homogeneous numeric and boolean arrays expose from_slice_*_array_into variants that reuse the result vector's capacity.

Consecutive owned items can be decoded from an input stream with ReaderDeserializer::deserialize_next; its byte offset is absolute within the stream. Refillable input cannot safely produce borrowed values, so use the slice Deserializer when output borrows strings or bytes.

Typed semantic tags use Tagged<T>:

let bytes = fcpw::to_vec(&fcpw::Tagged::with_tag(32, "https://example.com"))?;
let tagged: fcpw::Tagged<String> = fcpw::from_slice(&bytes)?;
assert_eq!(tagged.tag, Some(32));
# Ok::<(), fcpw::Error>(())

into_writer(value, writer) provides Ciborium's argument order, while de, ser, tag, and value modules ease imports during migration. write_self_describe emits tag 55799. For serde-cbor's packed representation, use to_vec_packed or EncodeConfig::packed(); packed encoding makes Rust field and variant declaration order part of the wire format.

The default feature set is std, alloc, and serde. The scalar parser remains available with --no-default-features. parallel enables Rayon-backed sequence and delimited-batch decoding, while diagnostic enables RFC 8949 diagnostic formatting and parsing.

API and features

• SliceDecoder, Parser, and Encoder provide allocation-free core CBOR processing.
• Value and BorrowedValue preserve CBOR-specific values, including tags, simple values, undefined, ordered maps with duplicate keys, and bignums.
• Deserializer and Serializer expose stateful Serde integration; the one-shot from_slice, to_vec, to_slice, and writer helpers cover common cases.
• ReaderDeserializer handles consecutive owned values from a stream, while SequenceDecoder exposes RFC 8742 sequences over slices.
• Deterministic, packed, diagnostic, and parallel operations are explicit so their wire-format or dependency tradeoffs remain visible at call sites.

The alloc feature adds owned values and collection helpers. serde adds the Serde APIs and implies alloc; std adds reader/writer APIs; diagnostic adds diagnostic notation; and parallel adds Rayon-backed batch decoding and implies the other public API features it needs. A complete parity-oriented walkthrough is available in examples/api_parity.rs.

The crate contains no unsafe code. Architecture-specific acceleration is only enabled when it can retain the scalar implementation as its semantic oracle.

Benchmarks

Criterion benchmarks compare FCPW with cbor4ii, serde_cbor, and ciborium, and minicbor. The latest local run produced the following Criterion time estimates; lower is better and the fastest result in each row is bold.

Operation	Input	FCPW	cbor4ii	serde_cbor	Ciborium	minicbor
Decode small record	503 B	295 ns	515 ns	444 ns	1.40 µs	501 ns
Decode medium record	7,210 B	5.92 µs	7.78 µs	6.82 µs	21.1 µs	7.10 µs
Encode small record	503 B	182 ns	357 ns	492 ns	471 ns	523 ns
Encode medium record	7,210 B	1.73 µs	2.03 µs	4.13 µs	3.91 µs	4.82 µs

These results were measured on August 13, 2026, with Rust 1.97.1 on an Intel Core Ultra 7 155H running Linux, using FCPW 0.1.3, cbor4ii 1.2.2, serde_cbor 0.11.2, Ciborium 0.2.2, minicbor 2.3.0 through minicbor-serde 0.7.1, and Criterion 0.8.2. All five implementations encode these workloads to the same size. Results vary by hardware and toolchain; run cargo bench --bench codec to measure the benches/codec.rs suite locally.

MSRV and licensing

FCPW uses Rust 2024 and requires Rust 1.88 or newer. It is available under your choice of the MIT or Apache-2.0 license.