Go experiments with platform-independent SIMD
Go is trying to expose SIMD as portable, readable Go rather than architecture-specific assembly, using compiler specialization and runtime dispatch to preserve performance where hardware support exists.
Logged at IST: 2026-09-25 19:01 IST
What it is: David Chase and Junyang Shao describe Go's experimental SIMD APIs: architecture-dependent archsimd, plus a portable simd package added across Go 1.26 and 1.27.
Gist: Before this work, Go programs mostly had to drop to Go assembly to use CPU SIMD instructions directly. The new API stack gives performance-sensitive code a way to express vector operations in Go, while accounting for the messy reality that SIMD varies by platform: amd64 has AVX widths and features, arm64 has NEON and SVE, wasm has fixed 128-bit vectors, and mask semantics differ across families.
The portable simd package takes a conservative route. It removes fixed vector sizes from the public type system, exposes size-agnostic vector types such as simd.Uint8s and simd.Float32s, and supports operations that can be mapped across amd64 AVX/AVX2/AVX512, arm64 NEON, wasm SIMD, or emulated when native support is missing. Code opts in with GOEXPERIMENT=simd, and GODEBUG=simd=... can force emulation or specific vector widths for testing.
The implementation is also interesting compiler work. Go rewrites functions that mention simd types into specialized versions for vector lengths such as 128, 256, 512, or emulation, then dispatches at runtime high enough in the call graph to avoid per-operation overhead. The design goal is portable code that is still close to assembly performance when the source operations match the hardware.
Newsletter angle: A serious attempt to make SIMD feel like ordinary portable Go: architecture-aware under the hood, readable enough for humans and LLMs, and explicit about the abstraction/performance tradeoff.