Experiments: adding SIMD helpers

[samber]

Introducing github.com/samber/lo/exp/simd package.

Go 1.26 brings SIMD instructions to Go.

The feature is still experimental, so i'm not adding this to the core package of lo.

# usage
import losimd "github.com/samber/lo/exp/simd"

result := losimd.SumInt8x64([]int8{1, 2, 3, ...})

Requirements:

• amd64 architecture
• AVX2 or AVX512 CPU
• build with GOEXPERIMENT=simd

I've added SumInt8 variants that route the implementation to the right architecture automatically. I would be interested by your opinion on this.

Should we support vaargs ? If yes, should losimd.SumInt8x64(arg1, arg2, arg3) returns a slice of 3 results or a single result?

@Jorropo @d-enk

Warning: The experimental nature of this sub-package implies that we might have breaking changes in the future.

TODO:

• Benchmarks
• Documentation: it is an experimental feature, and we might break the API in the future

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.05%. Comparing base (3e269f8) to head (72c24d8).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #801   +/-   ##
=======================================
  Coverage   92.05%   92.05%           
=======================================
  Files          32       32           
  Lines        4205     4205           
=======================================
  Hits         3871     3871           
  Misses        255      255           
  Partials       79       79           

Flag	Coverage Δ
unittests	92.05% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

This PR introduces experimental SIMD (Single Instruction Multiple Data) support for Go 1.26, adding a new github.com/samber/lo/exp/simd package with optimized mathematical and search operations using SSE, AVX2, and AVX-512 instruction sets.

Changes:

• SIMD-accelerated math operations (Sum, Mean, Clamp, Min, Max) for all numeric types
• SIMD-accelerated Contains operations for searching
• Automatic CPU feature detection and dispatching to optimal implementation
• Comprehensive test coverage for all SIMD variants

Reviewed changes

Copilot reviewed 17 out of 18 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
exp/simd/math.go	Dispatcher functions that route to optimal SIMD implementation
exp/simd/math_sse.go	SSE (128-bit) SIMD implementations
exp/simd/math_avx2.go	AVX2 (256-bit) SIMD implementations
exp/simd/math_avx512.go	AVX-512 (512-bit) SIMD implementations
exp/simd/intersect_*.go	SIMD Contains implementations for SSE/AVX2/AVX-512
exp/simd/*_test.go	Comprehensive test files for all SIMD variants
exp/simd/cpu_amd64_test.go	CPU feature detection helpers for tests
exp/simd/README.md	Documentation on CPU compatibility and usage
exp/simd/go.mod	Module definition with Go 1.26 requirement

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[d-enk]

Wrong requireAVX512 for sse

[samber]

It uses mask.ToBits() which requires avx512.

We could imagine 2 conditional branches: AVX512 and non-AVX512?

[d-enk]

I didn't really look into it

The test just worked successfully for me.

[d-enk]

Only TestContainsUint16x8, TestContainsUint16x8 not works by illegal instruction

[samber]

Can you copy your /proc/cpuinfo please?

Since I code on MacOS, I need to start a VM to test all architectures. 🥵

Having AVX512 instructions but not AVX2 seems weired.

[d-enk]

archsimd.X86Features

HasAES: true
HasADX: true
HasAVX: true
HasAVXVNNI: false
HasAVX2: true
HasAVX512: false
HasAVX512F: false
HasAVX512CD: false
HasAVX512BW: false
HasAVX512DQ: false
HasAVX512VL: false
HasAVX512GFNI: false
HasAVX512VAES: false
HasAVX512VNNI: false
HasAVX512VBMI: false
HasAVX512VBMI2: false
HasAVX512BITALG: false
HasAVX512VPOPCNTDQ: false
HasAVX512VPCLMULQDQ: false
HasBMI1: true
HasBMI2: true
HasERMS: true
HasFSRM: false
HasFMA: true
HasGFNI: false
HasOSXSAVE: true
HasPCLMULQDQ: true
HasPOPCNT: true
HasRDTSCP: true
HasSHA: false
HasSSE3: true
HasSSSE3: true
HasSSE41: true
HasSSE42: true
HasVAES: false

[samber]

i'm pretty sure to have fixed all architecture checks

Can you run tests again please?

For CI, I moved to ubicloud, which supports AVX512. Github seems to have multiple hardware configs and none supports AVX512 during the weekend :-(

[d-enk]

You can also use for optimization:

• const lanes
• uint index

For example, it became ~30% faster.

func SumInt16x8[T ~int16](collection []T) T {
	length := uint(len(collection))
	if length == 0 {
		return 0
	}

	const lanes = 8

	base := unsafeSliceInt16(collection, int(length))
	var acc archsimd.Int16x8

	i := uint(0)
	for ; i+lanes <= length; i += lanes {
		v := archsimd.LoadInt16x8Slice(base[i : i+lanes])
		acc = acc.Add(v)
	}

	var buf [lanes]int16
	acc.Store(&buf)
	var sum T
	for k := range uint(lanes) {
		sum += T(buf[k])
	}

	for ; i < length; i++ {
		sum += collection[i]
	}

	return sum
}