internal/cpu: repair VNNI feature check

This is a pain to test.
Also the original test was never executed, because it was wrong.

It looks like processors that might lack this features
include Intel 11th generation and AMD Zen 4.  These might
or might not have bit 2 set in the 7th cpuid "leaf" (SM4)
which is what the incorrect test was checking; the bug
is triggered by ^VNNI & SM4.  Apparently the SM4 bit is
not usually set, else we would have seen a test failure.

The "Lion Cove" microarchitecture (Arrow Lake, Lunar Lake)
appears to trigger this problem, it's not clear if there are
others.  It was hard to verify this from online information.

Fixes #76881.

Change-Id: I21be6b4f47134d81e89799b0f06f89fcb6563264
Reviewed-on: https://go-review.googlesource.com/c/go/+/731240
TryBot-Bypass: David Chase <drchase@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>

diff --git a/src/internal/cpu/cpu_x86.go b/src/internal/cpu/cpu_x86.go
index 4610ce807eea6a197110dd0cd58bd5b40997c0f6..711fb045c3370b74e7b07a46ee5af16286b163be 100644 (file)
--- a/src/internal/cpu/cpu_x86.go
+++ b/src/internal/cpu/cpu_x86.go
@@ -219,7 +219,7 @@ func doinit() {
        if eax7 >= 1 {
                eax71, _, _, _ := cpuid(7, 1)
                if X86.HasAVX {
-                       X86.HasAVXVNNI = isSet(4, eax71)
+                       X86.HasAVXVNNI = isSet(eax71, cpuid_AVXVNNI)
                }
        }
 

diff --git a/src/simd/archsimd/internal/simd_test/simd_test.go b/src/simd/archsimd/internal/simd_test/simd_test.go
index 83925ae789b0f008c93dc267a468edfd1e7d8fbf..5fd7407dbee0c0a566ef7e24dab85b9de464dd14 100644 (file)
--- a/src/simd/archsimd/internal/simd_test/simd_test.go
+++ b/src/simd/archsimd/internal/simd_test/simd_test.go
@@ -1135,18 +1135,22 @@ func TestDotProductQuadruple(t *testing.T) {
        wanted2 := make([]int32, 4)
        res1 := make([]int32, 4)
        res2 := make([]int32, 4)
+       for i := range 16 {
+               xd[i] = int8(i + 112)  // 112+15 = 127
+               yd[i] = uint8(i + 240) // 240+15 = 255
+       }
        for i := range 4 {
-               xd[i] = 5
-               yd[i] = 6
-               zd[i] = 3
-               wanted1[i] = 30
-               wanted2[i] = 30
+               i4 := 4 * i
+               wanted1[i] = int32(xd[i4])*int32(yd[i4]) + int32(xd[i4+1])*int32(yd[i4+1]) + int32(xd[i4+2])*int32(yd[i4+2]) + int32(xd[i4+3])*int32(yd[i4+3])
+               zd[i] = int32(i + 1)
+               wanted2[i] = wanted1[i] + zd[i]
        }
+
        x := archsimd.LoadInt8x16Slice(xd)
        y := archsimd.LoadUint8x16Slice(yd)
        z := archsimd.LoadInt32x4Slice(zd)
        x.DotProductQuadruple(y).StoreSlice(res1)
-       x.DotProductQuadruple(y).Add(z).StoreSlice(res1)
+       x.DotProductQuadruple(y).Add(z).StoreSlice(res2)
        for i := range 4 {
                if res1[i] != wanted1[i] {
                        t.Errorf("got %d wanted %d", res1[i], wanted1[i])