Expand description
Runtime SIMD dispatch for the Wickra technical indicators core.
Release binaries are built for the baseline of their target — on x86_64
that is SSE2, with no AVX and no fused multiply-add instruction — so they
run on every CPU of the architecture. A batch kernel that would profit from
AVX2 lanes or a hardware FMA therefore cannot simply be compiled for them.
dispatch closes that gap: it checks the CPU once (the standard library
caches the answer) and runs a Kernel inside a function compiled with
AVX2 and FMA enabled when both are present, and in the baseline build
otherwise. A kernel that opts in (Kernel::WIDE) runs with AVX-512F
enabled as well where the CPU has it, and does its element-wise work in
eight-lane Simd::W values: one 512-bit register there, two four-lane
vectors everywhere else.
§Same result on every path
Dispatch changes how a kernel is compiled, never what it computes. Rust
evaluates floating-point code exactly as written — it never reassociates,
and never contracts a * b + c into a fused multiply-add — so the AVX2
build and the baseline build of one kernel perform the same IEEE-754
operations in the same order. f64::mul_add is correctly rounded whether it
becomes the FMA instruction or the C library’s fma, and an eight-lane
operation is the same IEEE-754 operation on each lane whether the lanes sit
in one register or two. Every path therefore returns bit-identical values;
the crate’s tests check exactly that, at every level the machine has.
The AVX-512 level needs Rust 1.89 or later, where its target feature is stable; built by an older compiler the crate has no such level, and a kernel that opts in runs at the AVX2 one – the same bits, at the AVX2 speed.
On aarch64, NEON and FMA are part of the baseline, so kernels already get
them and dispatch runs them directly.
§Writing a kernel
Implement Kernel and mark Kernel::run #[inline(always)], as well
as any helper it calls: code is only compiled with the enabled features if
it is inlined into the dispatching function.
Structs§
- Avx2
- Proof that this CPU implements AVX2 and FMA; only
crate::dispatchmakes one. - Avx512
- Proof that this CPU implements AVX-512F (and so AVX2 and FMA); only
crate::dispatchmakes one. - Portable
- Lanes as
[f64; 4]in ordinary Rust arithmetic.
Enums§
Traits§
- Kernel
- A unit of work
dispatchcan run with the best instruction set available. - Simd
- Four
f64lanes and the operations the Wickra kernels are built from.
Functions§
- dispatch
- Run
kernelwith the best instruction set this CPU offers. - level
- The instruction set
dispatchruns kernels with on this machine. - run_
baseline - Run
kernelin the baseline build with thePortablelanes, never dispatching. Exists so tests (in this crate and in its users) can prove a kernel returns the same bits on every path.