Expand description
The sole crate permitted to contain audited kernel unsafe islands.
Every future unsafe kernel must be feature-gated, carry a SAFETY: comment,
and retain a bit-identical safe scalar fallback.
§Frankentorch boundary
This Phase-0 scaffold intentionally has no frankentorch dependency: it does not yet expose a kernel facade that consumes one. Add each dependency only alongside a real facade API and its scalar fallback; do not predeclare a machine-local or otherwise unused substrate dependency.
§Permanent integer-kernel law
A route cannot become dispatchable until selftest::run_selftest proves its all-extreme
reduction at every model-specific census binding row against an independent i64 oracle: the
U8S8 envelope (255 * 127 * K, the ceiling for a future unsigned-activation VNNI route) on
the checked scalar path, and the S8S8 kernel contract (±127 * 127 * K) through the real
int8::dot_i32 on every available tier with scalar equality. Every native tier, Q4 unpack
path, codec convolution, verifier, and batched variant extends that same surface before it
may be selected.
Modules§
- enhance
- FastEnhancer-S 48 kHz speech denoiser — a pure-Rust port of the pinned reference.
- f32ref
- f32 reference kernels: the correctness baseline every optimized tier must reproduce.
- int4
- W4A8 int4 weights for the microdecoder — packed two-per-byte, unpacked in registers.
- int8
- Int8 W8A8 kernels: symmetric per-output-channel Q8 weights times per-row Q8 activations.
- mmap
- Audited OS-interface island: read-only memory mapping of a checkpoint file.
- packed_
gemm - Register-tiled, panel-packed f32 GEMM — the BLAS-shaped dense route for hosts with no BLAS.
- route
- The one process-wide switch between the optimized product route and the f32 reference.
- selftest
- Shipped integer-overflow proofs for the kernels this crate can dispatch.
- sleef
- SLEEF’s 1-ulp
sinfandexpf, ported to safe scalar Rust. - startup_
env - Audited startup-time environment defaulting.
- team
- KernelTeam v0: the persistent worker pool for int8 GEMV/GEMM output-column partitions.
Constants§
- SCAFFOLD_
REVISION - Identifies this crate’s scaffold revision.