simd/archsimd: X86.AVX512() always false on darwin/amd64

[vsivsi]

Go version

go version go1.26.2 darwin/amd64

Output of go env in your module/workspace:

go env output (click to expand)

AR='ar'
CC='cc'
CGO_CFLAGS='-O2 -g'
CGO_CPPFLAGS=''
CGO_CXXFLAGS='-O2 -g'
CGO_ENABLED='1'
CGO_FFLAGS='-O2 -g'
CGO_LDFLAGS='-O2 -g'
CXX='c++'
GCCGO='gccgo'
GO111MODULE='on'
GOAMD64='v1'
GOARCH='amd64'
GOAUTH='netrc'
GOBIN=''
GOCACHE='/Users//Library/Caches/go-build'
GOCACHEPROG=''
GODEBUG=''
GOENV='/Users//Library/Application Support/go/env'
GOEXE=''
GOEXPERIMENT=''
GOFIPS140='off'
GOFLAGS=''
GOGCCFLAGS='-fPIC -arch x86_64 -m64 -pthread -fno-caret-diagnostics -Qunused-arguments -fmessage-length=0 -ffile-prefix-map=/var/folders/cb/2vx9ptwx6qq3ghc6jgcsr1m80000gn/T/go-build3487313098=/tmp/go-build -gno-record-gcc-switches -fno-common'
GOHOSTARCH='amd64'
GOHOSTOS='darwin'
GOINSECURE=''
GOMOD='/Users//avx512test/go.mod'
GOMODCACHE='/Users//go/pkg/mod'
GONOPROXY='github.com/vsivsi'
GONOSUMDB='github.com/vsivsi'
GOOS='darwin'
GOPATH='/Users//go'
GOPRIVATE='github.com/vsivsi'
GOPROXY='https://proxy.golang.org,direct'
GOROOT='/usr/local/Cellar/go/1.26.2/libexec'
GOSUMDB='sum.golang.org'
GOTELEMETRY='local'
GOTELEMETRYDIR='/Users//Library/Application Support/go/telemetry'
GOTMPDIR=''
GOTOOLCHAIN='auto'
GOTOOLDIR='/usr/local/Cellar/go/1.26.2/libexec/pkg/tool/darwin_amd64'
GOVCS=''
GOVERSION='go1.26.2'
GOWORK=''
PKG_CONFIG='pkg-config'

What did you do?

On an Intel iMac Pro with an Intel(R) Xeon(R) W-2150B CPU @ 3.00GHz processor running MacOS version 15.7.5 (24G624), I attempted to use the new simd/archsimd experiment to implement a test function using AVX-512 intrinsics. The first step is to ensure the machine at runtime supports AVX-512 using the provided archsimd.X86.AVX512() call, which should always return true on a machine with AVX-512.

The following file tests whether that condition is met by comparing output with the golang.org/x/sys/cpu package cpu.X86.HasAVX512 flag.

//go:build goexperiment.simd && amd64

package archsimd_parity_test

import (
        "testing"

        "simd/archsimd"
        "golang.org/x/sys/cpu"
)

// TestAVX512Parity verifies that archsimd.X86.AVX512() and
// cpu.X86.HasAVX512 always agree. The two APIs are expected to reflect
// the same underlying CPU capability; any divergence would indicate a
// bug in one of the implementations.
func TestAVX512Parity(t *testing.T) {
        archsimdResult := archsimd.X86.AVX512()
        sysResult := cpu.X86.HasAVX512

        if archsimdResult != sysResult {
                t.Fatalf(
                        "AVX512 parity check FAILED: "+
                                "archsimd.X86.AVX512() = %v, "+
                                "cpu.X86.HasAVX512 = %v. "+
                                "\nThese two values should always be equal. ",
                        archsimdResult, sysResult,
                )
        }

        t.Logf("AVX512 parity OK: archsimd.X86.AVX512() == cpu.X86.HasAVX512 == %v", archsimdResult)
}

What did you see happen?

My test machine (iMac Pro) supports AVX-512, as reported by the darwin sysctl command:

echo "macOS $(sw_vers -productVersion) — $(sysctl -n machdep.cpu.brand_string)" && \
echo "AVX-512 features present:" && \
sysctl machdep.cpu.leaf7_features 2>/dev/null | \
     grep -o 'AVX512[A-Z]*' || \
echo "No AVX-512 support detected"

Outputs:

macOS 15.7.5 — Intel(R) Xeon(R) W-2150B CPU @ 3.00GHz
AVX-512 features present:
AVX512F
AVX512DQ
AVX512CD
AVX512BW
AVX512VL

Running the test script from the previous section:

Save the file above in an new directory as e.g. avx512_test.go. Then:

go mod init avx512-parity
go get golang.org/x/sys/cpu
GOEXPERIMENT=simd go test -v -run TestAVX512Parity .

Outputs:

go: creating new go.mod: module avx512-parity
go: to add module requirements and sums:
	go mod tidy
go: added golang.org/x/sys v0.43.0
=== RUN   TestAVX512Parity
    avx512_test.go:21: AVX512 parity check FAILED: archsimd.X86.AVX512() = false, cpu.X86.HasAVX512 = true. 
        These two values should always be equal. 
--- FAIL: TestAVX512Parity (0.00s)
FAIL
FAIL	avx512-parity	0.270s
FAIL

Repeating the above tests on my MacBook Pro:

macOS 26.4 — Intel(R) Core(TM) i7-1068NG7 CPU @ 2.30GHz
AVX-512 features present:
AVX512F
AVX512DQ
AVX512IFMA
AVX512CD
AVX512BW
AVX512VL
AVX512VBMI
AVX512VNNI
AVX512BITALG
AVX512VPOPCNTDQ

And:

go: creating new go.mod: module avx512-parity
go: to add module requirements and sums:
	go mod tidy
go: added golang.org/x/sys v0.43.0
=== RUN   TestAVX512Parity
    avx512_test.go:21: AVX512 parity check FAILED: archsimd.X86.AVX512() = false, cpu.X86.HasAVX512 = true. 
        These two values should always be equal. 
--- FAIL: TestAVX512Parity (0.00s)
FAIL
FAIL	avx512-parity	0.360s
FAIL

What did you expect to see?

I expect archsimd.X86.AVX512() return value to reflect the true AVX-512 capabilities of the hardware, and to be consistent with the output of cpu.X86.HasAVX512 from golang.org/x/sys/cpu.

[gabyhelp]

Related Issues

• x/sys/cpu: cpu.X86.HasAVX512 is incorrectly always false on darwin #43089 (closed)
• x/sys/cpu: ARM64.HasAES = false on Apple Silicon (M1 Max) #68150 (closed)
• x/sys/cpu: cpu.ARM64.SHA512 not correctly set for Mac M1 #76221 (closed)
• internal/cpu: the VNNI feature check is incorrect #76881 (closed)
• internal/cpu: should detect AVX512 #26505 (closed)
• x/sys/cpu: add support for detecting AVX-512 #41288 (closed)
• internal/cpu: cpu feature detect failed on darwin/arm64 #24753 (closed)
• simd/archsimd: archsimd.LoadUint8x32Slice() emits nops #76968 (closed)

Related Code Changes

• internal/cpu: detect support of AVX512

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[vsivsi]

Reproducing here the salient parts from a previous discussion.

The underlying issue seems to be that the archsimd.X86Features struct is initialized using imports from internal/cpu.

See: https://cs.opensource.google/go/go/+/refs/tags/go1.26.2:src/simd/archsimd/cpu.go;l=7-9

internal/cpu hasn't been updated with the correct darwin AVX512 detection support implemented in x/sys/cpu.

Specifically, this comment and its implications are no longer valid: https://cs.opensource.google/go/go/+/refs/tags/go1.26.2:src/internal/cpu/cpu_x86.go;l=147-151

The referenced issue in the highlighted comment in the above internal/cpu file has been fully resolved and released over 18 months ago: https://go-review.googlesource.com/c/sys/+/620256

Compare the x/sys/cpu implementation to the internal/cpu to understand why one works correctly in this case and the other doesn't: https://cs.opensource.google/go/x/sys/+/master:cpu/cpu_x86.go;l=171-177

A quick workaround for archsimd could be to import from x/sys/cpu rather than internal/cpu (since they implement the same API w.r.t. detecting amd64 simd features), although that may not be desirable for a core standard library. The permanent solution is probably for internal/cpu to incorporate the correct darwin/amd64 AVX-512 detection process already implemented in x/sys/cpu.

[cagedmantis]

cc @drchase @JunyangShao