sklears-simd

Latest release: 0.2.0 (July 14, 2026). See the workspace release notes for highlights and upgrade guidance.

Overview

sklears-simd exposes low-level SIMD, GPU, and hardware acceleration utilities used across the sklears ecosystem. While primarily an internal crate, it is documented for contributors building new high-performance components.

Key Features

• Vector Abstractions: Portable SIMD types (F32x4, F32x8, F32x16) with architecture-specific intrinsics (SSE2, AVX2, AVX512, NEON) and scalar fallbacks.
• Alignment & Memory: Alignment helpers, prefetching hints, cache-aware allocation strategies.
• Accelerator Scaffolding: Experimental FPGA/TPU/quantum/neuromorphic interface abstractions with CPU SIMD fallback (no vendor hardware backend). Real GPU dispatch lives in sklears-core's oxicuda-backed gpu module, not in this crate.
• Benchmark Harnesses: Criterion-based benchmarks and profiling utilities for micro-optimizations.

Quick Peek

use sklears_simd::vector::F32x4;

let a = F32x4::new(1.0, 2.0, 3.0, 4.0);
let b = F32x4::splat(2.0);
let result = a.mul(b);
assert_eq!(result.horizontal_sum(), 20.0);

Status

• Core building block for 0.2.0; 518 passing crate tests (4 skipped, --features parallel).
• Powers CPU SIMD acceleration in linear models, neighbors, metrics, and more; GPU dispatch is handled by sklears-core.
• Contributor roadmap (new architectures, auto-vectorization tooling) maintained in TODO.md.