Examples of methods to use SIMD in Go

Demonstrated is an example which calculates the sum of two slices of float32 using SIMD instructions.

result[i] = a[i] + b[i]

The example is implemented in three different ways:

• code generated Assembler by avo (only x86 possible)
• compile and translate to Go Assembler by gocc (any architecture supported by clang)
• call SIMD instructions directly in Go code with go-simd (only ARM possible)

how to install gocc

To install see gocc README

other notable examples

See kelindar/bitmap for an example using auto-vectorization in combination with gocc.

Benchmark results

x86 (on AMD Ryzen 7 5700X)

goos: linux
goarch: amd64
pkg: github.com/Crash129/go-simd-example
cpu: AMD Ryzen 7 5700X 8-Core Processor
BenchmarkAvo
BenchmarkAvo-16        	 3100104	       383.8 ns/op	       0 B/op	       0 allocs/op
BenchmarkGocc
BenchmarkGocc-16       	 3770598	       318.4 ns/op	       0 B/op	       0 allocs/op
BenchmarkPlainGo
BenchmarkPlainGo-16    	  666050	      1797 ns/op	       0 B/op	       0 allocs/op
BenchmarkGoSIMD
    main_test.go:49: no ARM SIMD support
--- SKIP: BenchmarkGoSIMD
PASS

Note that the assembler code of the avo implementation has no optimizations. avo is able to have the same performance as gocc but the generated assembler code of gocc is already optimized by clang.

ARM (on Apple M2 Pro)

goos: darwin
goarch: arm64
pkg: github.com/Crash129/go-simd-example
BenchmarkAvo
    main_test.go:28: no AVX2 support
--- SKIP: BenchmarkAvo
BenchmarkGocc
BenchmarkGocc-12        15680464               374.5 ns/op
BenchmarkPlainGo
BenchmarkPlainGo-12      3339758              1798 ns/op
BenchmarkGoSIMD
BenchmarkGoSIMD-12       4593692              1309 ns/op
PASS