mirror of
https://github.com/luxfi/p3q.git
synced 2026-10-30 15:22:47 +00:00
zeekay
08a45df14c
Adds a Sha3Backend trait in p3q-sha3 with CPU + GPU impls. The CPU backend is the audited FIPS 202 reference (sha3 crate); the GPU backend (new crate p3q-gpu-sha3) dlopens libluxgpu and dispatches through lux_gpu_sha3_256_batch / lux_gpu_shake256_batch. Transparent CPU fallback whenever dlopen fails, GPU context creation fails, or libluxgpu panics inside FFI (catch_unwind boundary). KAT parity enforced against FIPS 202 reference vectors. The verifier opts in via --features gpu-sha3. Per-query Merkle authentications fan out into one verify_proof_batch per Merkle root (trace + each FRI layer); the backend gets one call per root rather than O(num_queries * log N) serial hashes. CpuShake overrides tuple_hash256_batch with a streaming impl that avoids per-input payload allocations — the CPU-only build is now 1017 us/iter (down from 1308 us serial baseline on the toy AIR). libluxgpu 0.4 exposes SHAKE256 (0x1F domain) but not cSHAKE256 (0x04), so the audited transcript binding (cSHAKE / TupleHash / KMAC) still runs on CPU. The wiring is in place: when libluxgpu adds a native cSHAKE batch the override drops into p3q-gpu-sha3 with zero verifier change. Hard constraints honoured: - p3q-sha3 stays #![forbid(unsafe_code)]; p3q-gpu-sha3 is the second `unsafe`-allowed crate (alongside p3q-c-abi), each unsafe block paired with catch_unwind - GpuShake field order is load-bearing: `gpu` drops before `lib` to avoid use-after-free on the destroy fn-ptr at shutdown - transcript bytes UNCHANGED — Sha3Backend contract guarantees byte-identical output on every backend - 101/101 workspace tests pass with and without --features gpu-sha3 Bench (toy AIR, 256-iter avg, three runs each): CPU-only: 1015.2 / 1018.6 / 1017.4 us/iter --features gpu-sha3: 1010.3 / 1020.0 / 1019.4 us/iter
28 lines
1.1 KiB
TOML
28 lines
1.1 KiB
TOML
[package]
|
|
name = "p3q-gpu-sha3"
|
|
description = "P3Q SHA3-256 / SHAKE256 GPU backend (libluxgpu 0.4 FFI). Optional plug-in for p3q-sha3 when --features gpu-sha3 is enabled."
|
|
version.workspace = true
|
|
authors.workspace = true
|
|
edition.workspace = true
|
|
license.workspace = true
|
|
repository.workspace = true
|
|
homepage.workspace = true
|
|
rust-version.workspace = true
|
|
keywords = ["sha3", "shake256", "gpu", "ffi", "post-quantum"]
|
|
categories = ["cryptography", "external-ffi-bindings"]
|
|
|
|
[dependencies]
|
|
p3q-sha3 = { workspace = true }
|
|
# dlopen-based FFI: the libluxgpu.dylib / .so is loaded at runtime, not
|
|
# linked at build time, so that p3q binaries stay portable across hosts
|
|
# that may or may not have the GPU library installed.
|
|
libloading = "0.8"
|
|
|
|
# This crate is the ONE place in the workspace where `unsafe` is
|
|
# unavoidable: dlsym'd C entry points return `*mut LuxGPU` opaque
|
|
# pointers and consume raw `*const u8` buffers. Every unsafe block is
|
|
# paired with a `catch_unwind` boundary so a libluxgpu panic NEVER
|
|
# crosses the FFI seam back into the audited verifier.
|
|
[lints.rust]
|
|
unsafe_code = "allow"
|