FP16 Hardware Feature Detection

Open
#11 1 comment 1 reaction 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
5/5
Estimated time
Over a week
Newbie friendliness
25/100
Issue type
Feature
Clarity
Mostly clear
Activity status
Active
Tech stack
wasm

Research direction

Start with the linked Lowering.md and compare the three host-detection options and open questions in this issue. The issue does not identify proposal files, tests, or a settled design; first determine the intended interface and how it fits WebAssembly semantics. Done means an agreed, specified way to report FP16 support tiers, including forward compatibility and the stated stability and privacy constraints.

Written by the indexing model from the issue text.

Description

Motivation

Host CPUs fall into three tiers of FP16 hardware support (see Lowering.md):

Tier Hardware support Examples
0 None. All FP16 ops are emulated in software x86 without F16C (e.g. Jasper Lake), RISC-V without Zfhmin/Zvfhmin, interpreters
1 f16 <-> f32 conversions only. Other f16x8 ops lower via wider floats or software Client x86-64 with F16C (AMD Zen 1–5, Intel Core through Arrow/Lunar Lake), AArch64 (ARMv8.0-A), RISC-V Zfhmin+Zvfhmin
2 Native f16x8 SIMD ARMv8.2+ FEAT_FP16, server x86 AVX512-FP16+VL / AVX10, RISC-V Zvfh

"Conversions" means f32.load_f16, f32.store_f16, f32x4.promote_low_f16x8, f16x8.demote_f32x4_zero, and f16x8.{splat,extract_lane,replace_lane} (which convert lanes to/from f32).

Every engine supporting FP16 validates all instructions on all three tiers, so a single .wasm binary runs everywhere and can bundle both f16x8 and f32x4 code paths. Because everything validates, WebAssembly.validate only reports engine support, not whether the host CPU accelerates FP16.

Libraries such as XNNPACK need the hardware tier to pick the right kernel:

  • Tier 2: compute directly in f16x8.
  • Tier 1: store weights in f16 (halving memory bandwidth), promote on load, and compute in f32x4. Running a pure f16x8 kernel on tier 1 (nearly all consumer x86-64 CPUs) is slower than f32x4 because each arithmetic op unpacks to wider floats, computes, and repacks (f16x8.madd/nmadd cannot even use f32 FMA without double rounding).
  • Tier 0: compute in f32x4 (unpacking any f16 weights once up front rather than in inner loops).

Without feature detection, the only way to choose at runtime is a microbenchmark at startup, which is slow, noisy, and adds to startup time.

Requirements

All options below share these properties:

  • Hint only: the result does not affect validation or instruction semantics. Engines may conservatively return 0 (e.g. in anti-fingerprinting/jitless modes or the deterministic profile).
  • Stable: the value is constant for at least a compiled module or instance lifetime.
  • Low fingerprinting surface: at most a couple of coarse bits, which are already observable via timing.

Options

Option 1: Dedicated instruction (f16.feat)

f16.feat : [] -> [i32] returns the tier from the table above.

  • Pros: Self-contained and works in any embedding (Web, WASI) without glue code. Engines can constant-fold it at compile time and eliminate dead inline branches. JS loaders (e.g. wasm-feature-detect) can still probe it before fetching large bundles via a tiny module.
  • Cons: Adds a host-dependent instruction to the core spec (though relaxed-simd sets precedent for embedder-chosen hardware behavior). Specific to FP16, so other proposals would need their own.
Option 2: Opcode-parameterized query (instr.support <opcode>)

instr.support <opcode> : [] -> [i32] returns 1 if the given opcode is natively accelerated on the host, 0 otherwise.

  • Pros: General. Also covers relaxed SIMD (e.g. FMA, dot product) and future proposals without new opcodes. Can be constant-folded like Option 1.
  • Cons: "Accelerated" is ill-defined per opcode. For example, f16x8.abs/neg are a single bitwise op (sign-bit mask) on any SIMD hardware, yet say nothing about whether f16x8.add is fast. Larger fingerprinting surface. Broadens the scope of the FP16 proposal. Encoding an opcode as an immediate will be awkward.
Option 3: JS API (WebAssembly.getHardwareSupport(name))
const tier = WebAssembly.getHardwareSupport("fp16"); // 0, 1, 2, or undefined if unsupported
  • Pros: Keeps host introspection out of the core spec. Extensible to other features via new names.
  • Cons: Web-only unless mirrored in WASI/C APIs. Pure-Wasm libraries can't self-detect, so the embedder must plumb the result in. Constant folding requires pairing with a compile-time import (e.g. via Wasm 3.0 builtins).

Open questions

  • Single bit vs. 3 tiers? A single bit ("is native f16x8 available?") merges tiers 0 and 1, even though tier 0 still occurs on recent low-power x86 (e.g. Jasper Lake Chromebooks), non-Zfhmin RISC-V, VMs, interpreters, and privacy modes. Without distinguishing them, libraries can't choose between the tier 1 promote-on-load path and the tier 0 unpack-up-front path above.
  • Ordered tier (0, 1, 2) vs. bitmask? An ordered tier requires callers to check tier >= 2 to stay forward-compatible, whereas a bitmask (1 << 0 = conversions, 1 << 1 = f16x8 SIMD, upper bits 0) also handles non-linear extensions (e.g. follow-on widening FP16->FP32 FMA such as ARM FEAT_FHM).
Dominant language
WebAssembly
Stars
14
Forks
0
PR merge metrics
No merged PRs in 30d

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from WebAssembly/fp16

All issues in WebAssembly/fp16

Similar issues

More Operating Systems issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.