rustlane

rustlane brings an ISPC-style SPMD programming model to Rust as a pure library — you write natural scalar-looking control flow (if, while, for, break, early return) over varying values, and proc macros lower it to masked SIMD over nightly std::simd. No compiler fork, no external toolchain: just #[kernel], #[export], and foreach!.

rustlane is an independent project, not affiliated with Intel.

Status

Experimental. Nightly-only. The API is unstable and will change without notice.

What a kernel looks like

The SPMD idea: write one program as if it runs on a single lane, and the library runs it across a whole SIMD vector of lanes at once. Control flow that diverges between lanes is handled for you. Here is the Mandelbrot escape-time kernel — plain for / if / break over varying values:

#![feature(portable_simd)]
use rustlane::prelude::*;
use rustlane::kernel;

#[kernel]
fn mandel(c_re: Varying<f32>, c_im: Varying<f32>, count: i32) -> Varying<i32> {
    let mut z_re = c_re;
    let mut z_im = c_im;
    let mut ret = Varying::splat(0);
    for i in 0..count {
        if z_re * z_re + z_im * z_im > 4.0 {
            break;
        }
        let new_re = z_re * z_re - z_im * z_im;
        let new_im = 2.0 * z_re * z_im;
        unmasked! {
            z_re = c_re + new_re;
            z_im = c_im + new_im;
        }
        ret = i + 1;
    }
    ret
}

The macro rewrites this into masked std::simd: the varying if/break become per-lane mask updates and blends (ret = i + 1 stores only into lanes still iterating), and the for loop keeps running until every lane's mask is off, so a lane that has already escaped stops contributing without stopping the others.

Quick start

Nightly only. Select the nightly toolchain and add rustlane to your dependencies:

# rust-toolchain.toml
[toolchain]
channel = "nightly"

# Cargo.toml
[dependencies]
rustlane = "0.1.0"

A #[kernel] is the per-lane program; an #[export] is the all-uniform entry point that drives it with foreach! and is callable as an ordinary safe Rust function:

#![feature(portable_simd)]
use rustlane::prelude::*;
use rustlane::{export, kernel};

#[kernel]
fn scale(x: Varying<f32>, factor: f32) -> Varying<f32> {
    x * factor
}

#[export]
fn scale_all(input: &[f32], output: &mut [f32], factor: f32) {
    foreach!(i in 0..input.len() {
        output[i] = scale(input[i], factor);
    });
}

fn main() {
    let input: Vec<f32> = (0..1024).map(|v| v as f32).collect();
    let mut output = vec![0.0f32; input.len()];
    scale_all(&input, &mut output, 2.0);
    assert_eq!(output[10], 20.0);
}

Performance

Seven kernels from ISPC's own example suite (options contributes two), measured against Intel ISPC on two machines. #[export] dispatches to the widest target the CPU supports — 8 lanes on the M2 Pro, 16 on the Zen4 box — and the ISPC target at that same width is the orange bar.

Geometric mean of rustlane / ISPC runtime at matched width:

Each round is 3 warm-up reps plus an internal min-of-15 (mandelbrot: 20); five rounds run interleaved over every binary in a fixed order with 2 s cool-downs. Bars are the mean over the five rounds, whiskers a 95% t interval. Every round validated its output against the ISPC reference. Full toolchain versions are printed under each chart; per-round raw timings land in results/<arch>.csv.

The C++ columns are the same source at -O3; scalar adds -fno-vectorize -fno-slp-vectorize.

make -C ispc-ref build-all       # ISPC + C++ baselines (needs ispc, clang++)
cargo build --release
./rustlane-bench/measure.sh      # ~25 min
python3 rustlane-bench/parse_measurements.py aarch64   # or x86_64
python3 rustlane-bench/make_charts.py aarch64          # or x86_64

What works

• if / else with divergent (varying) conditions
• while (varying condition) / for / bare loop, with break / continue / early return under divergence
• unmasked! { .. } — an all-lanes block, for loop-carried locals
• cif! / cwhile! — opt-in coherent control flow (ISPC-style any() guards)
• foreach! / foreach_2d! / foreach_tiled! — inline iteration, no closures
• #[derive(SpmdValue)] — a struct vectorizes to a generated SoA type VaryingS<N>; #[spmd(uniform)] keeps a field scalar
• math — ISPC-ported transcendentals (exp / log / pow / sin / cos / rsqrt / rcp)
• rng — LFSR113 combined-Tausworthe varying RNG
• reduce — horizontal reductions and cross-lane ops (broadcast / rotate / shift / scan / pack)
• Runtime target dispatch: SSE2 / SSE4.1 / AVX2 / AVX-512 on x86-64, NEON on aarch64

Nightly is required and will stay required: Varying<T, N> is built on #![feature(portable_simd)], which is unstable upstream.

Not planned

• Varying<S> for a struct S. Varying<T, N> is repr(transparent) over Simd<T, N>, so T has to be a scalar. Structs go through #[derive(SpmdValue)] to SoA instead.
• match on a varying scrutinee. Bindings, guards and exhaustiveness have no lane-wise meaning, and anything that does have one is an if / else chain written out. Use that.
• Explicit generic arguments on a kernel call (k::<8>(x)). The macro supplies ::<N, _> itself.

Static safety

The dangerous SPMD mistakes are compile errors, not silent races. Assigning to a uniform (scalar) variable while lanes are diverging would make every lane race on one location, so there is simply no such trait impl — the write fails to type-check with a domain-specific diagnostic:

error[E0277]: cannot assign to a value of type `i32` under execution context `VMask<N>`
  --> tests/ui/uniform_assign.rs:14:7
   |
14 |     s = 5;
   |       ^ this assignment target cannot be written under the current control-flow mask
   |
   = note: assigning to a uniform (scalar) variable under VARYING control flow is
     not supported: every lane would race on one location. Make the variable a
     `Varying`, hoist the assignment out of the varying branch, or wrap it in
     `unmasked!` if all-lanes semantics are intended

Two safety properties hold by construction:

• Inactive lanes never touch memory. Every masked gather/scatter carries the active mask as its hardware enable, so an out-of-bounds index on a masked-off lane neither faults nor affects results.
• 48-case compile-fail diagnostics suite. The static rejection rules for #[kernel] and #[export] are pinned by 48 checked-in trybuild snapshots — each asserts the error span points at the offending user token — plus an additional case covering the SpmdValue derive.

How it works

Each execution context (all-on, a uniform branch bool, a varying mask, a uniform branch under a varying mask) is a distinct Rust type that the macro threads through the kernel as a hidden first parameter; every operator, assignment, and control-flow construct resolves against that type by trait, so the macro itself never inspects a value or a type. Because the context is a type and the "is this uniform?" test is a const, monomorphization deletes the mask machinery on every path that turns out uniform: a kernel entered all-on compiles down to plain scalar branches and contiguous vector loads, with blends emitted only where control flow is genuinely divergent. There are no unconditional coherence guards — the one any() reduction per loop is the loop exit check, not a per-if tax. The whole kernel tree is #[inline(always)], so #[export] can stamp it out once per SIMD target inside a #[target_feature] shim and pick the widest at runtime with a single cached indirect call.

License

Licensed under the MIT License.

Portions of the benchmark kernels and math routines are ported from Intel's ISPC examples under their original BSD-3-Clause terms; see THIRD-PARTY.md for the full attributions.