mirror of
https://github.com/lux-apps/papers.git
synced 2026-10-30 18:15:21 +00:00
zeekay
98343d238e
Quasar Horizon (consensus composition over Pulsar / ML-DSA / BLS lanes
under Prism + LSS lifecycle), Warp 2.0 (Pulse-backed cross-chain envelope
with forward-only v1->v2 migration and lattigo ReadUint64Slice DoS fix),
benchmark report (microbench, reshare, activation, Warp 2.0 verify, Prism
composition, GPU headroom, cross-WAN latency), and NIST IR 8214C-style
submission package (spec, parameter sets, domain separation, serialization,
reference-impl pointers, evaluation report, security analysis, side-channel
considerations).
All four deliverables compile cleanly via pdflatex+bibtex and reference
the canonical proof bucket files at proofs/{definitions,pulsar,lss,quasar}.
No duplication of proof bodies; cite by file path.
71 lines
4.4 KiB
TeX
71 lines
4.4 KiB
TeX
\section{Architecture}
|
|
\label{sec:architecture}
|
|
|
|
LP-137 organizes every algorithm under a uniform directory layout in \texttt{luxcpp/crypto}:
|
|
|
|
\begin{lstlisting}[language=bash,caption={Per-algorithm directory layout}]
|
|
luxcpp/crypto/<alg>/
|
|
cpp/ # First-party C++23 CPU body
|
|
c-abi/ # C ABI shim (header + .cc)
|
|
gpu/
|
|
metal/ # First-party .metal kernel
|
|
cuda/ # First-party .cu kernel
|
|
wgsl/ # First-party .wgsl kernel
|
|
test/
|
|
cmake/ # FetchContent integration of test oracles
|
|
kat_*.json # Known-answer test vectors
|
|
<alg>_test.cpp # CPU canonical correctness test
|
|
<alg>_metal_test.cpp # CPU<->GPU byte-equality test
|
|
<alg>_cuda_test.cu
|
|
<alg>_wgpu_test.cpp
|
|
CMakeLists.txt # Builds lib<alg>.a, no vendor link
|
|
\end{lstlisting}
|
|
|
|
\subsection{Three layers, one byte-equality contract}
|
|
|
|
\begin{figure}[H]
|
|
\centering
|
|
\begin{tikzpicture}[
|
|
node distance=0.9cm,
|
|
every node/.style={font=\small},
|
|
box/.style={rectangle, draw, rounded corners=2pt, minimum width=3.4cm, minimum height=0.8cm, align=center},
|
|
byte/.style={->, thick, dashed},
|
|
]
|
|
\node[box] (ref) at (0,0) {Reference (Go)\\\texttt{gnark-crypto}, \texttt{blst}};
|
|
\node[box, below=of ref] (cpu) {CPU canonical\\(Lux first-party C++23)};
|
|
\node[box, below left=of cpu, xshift=-1cm] (metal) {Metal\\(first-party)};
|
|
\node[box, below=of cpu] (cuda) {CUDA\\(first-party)};
|
|
\node[box, below right=of cpu, xshift=1cm] (wgsl) {WGSL\\(first-party)};
|
|
\draw[byte] (ref) -- node[right] {byte-equal} (cpu);
|
|
\draw[byte] (cpu) -- node[left] {byte-equal} (metal);
|
|
\draw[byte] (cpu) -- node[right] {byte-equal} (cuda);
|
|
\draw[byte] (cpu) -- node[right] {byte-equal} (wgsl);
|
|
\end{tikzpicture}
|
|
\caption{Three-hop byte-equality contract. The CPU canonical is byte-equal to the Go reference; each GPU backend is byte-equal to the CPU canonical; transitively, every backend is byte-equal to the Go reference. One contract, three hops, every backend coupled to the same ground truth.}
|
|
\label{fig:byte-equality}
|
|
\end{figure}
|
|
|
|
\paragraph{Why CPU first-party is the canonical.} If the CPU body came from the same algorithm family as the test oracle (e.g.\ both vendored from \texttt{gnark-crypto}, or both from \texttt{blst}), the oracle would lose its adversarial-distance property: a bug in the upstream would propagate through both sides of the byte-equality check. First-party CPU plus a vendored audited test-oracle is the correct adversarial structure; a bug on either side fails the test. The vendor-as-canonical alternative was specifically rejected on 2026-04-27 (LP-137 \texttt{Rejected proposals}).
|
|
|
|
\subsection{Build topology}
|
|
|
|
C++ deps come exclusively via CMake \texttt{FetchContent} from the \texttt{luxcpp/*} fork organization. Conan was rejected as a build dependency. Forks: \texttt{luxcpp/intx}, \texttt{luxcpp/evmmax}, \texttt{luxcpp/blst}, \texttt{luxcpp/pqclean}, \texttt{luxcpp/ed25519-donna}, \texttt{luxcpp/blake3-reference}. All track upstream verbatim with strict semver tags (no prerelease suffixes; no \texttt{-luxfi}, no \texttt{-alpha}); divergence (if ever needed) signaled by next-minor bump.
|
|
|
|
\subsection{C-ABI surface}
|
|
|
|
The \texttt{c-abi/} layer exposes a minimal, brand-neutral C entry point per algorithm. Public symbols use the algorithm name verbatim:
|
|
|
|
\begin{lstlisting}[language=C,caption={Sample C-ABI signature (BN254 / EIP-196 + EIP-197 \cite{eip196,eip197})}]
|
|
int bn254_g1_add(uint8_t out[64], const uint8_t a[64], const uint8_t b[64]);
|
|
int bn254_g1_scalar_mul(uint8_t out[64], const uint8_t p[64],
|
|
const uint8_t s[32]);
|
|
int bn254_pairing(uint8_t out[384], const uint8_t* g1s, const uint8_t* g2s,
|
|
size_t k);
|
|
\end{lstlisting}
|
|
|
|
\noindent No \texttt{LUX\_*}, \texttt{LUXFI\_*}, or \texttt{HANZO\_*} prefix appears anywhere in the public surface (LP-137 \texttt{Hard rules}). Where an entry point is unimplemented, it returns the documented sentinel \texttt{CRYPTO\_ERR\_NOTIMPL}; production paths that depend on it must check.
|
|
|
|
\subsection{Test-oracle isolation}
|
|
|
|
The vendored upstream library is included via \texttt{FetchContent} in \texttt{<alg>/test/cmake/<oracle>.cmake} and linked into the test binary only. The shipped \texttt{lib<alg>.a} library never depends on it. Verification: the production target's \texttt{ldd} output must contain none of \texttt{libblst.so}, \texttt{libgmp.so} (modexp non-test path), \texttt{libcrypto.so} (BoringSSL), etc. CI enforces this with a per-target audit step.
|