Basic chore cleanups of tests and benchmarks (#66)

* Upgrade cpuid, regen buildtags

* Move all uses of cpuid to one place, adjust benchmarks

* Rename SIMD functions to match instruction set

* One more rename, align arch-specific files

cpuid_other.go

@@ -23,6 +23,11 @@ import (
 	"github.com/klauspost/cpuid/v2"
 )
 
+var (
+	hasIntelSha = runtime.GOARCH == "amd64" && cpuid.CPU.Supports(cpuid.SHA, cpuid.SSSE3, cpuid.SSE4)
+	hasAvx512   = cpuid.CPU.Supports(cpuid.AVX512F, cpuid.AVX512DQ, cpuid.AVX512BW, cpuid.AVX512VL)
+)
+
 func hasArmSha2() bool {
 	if cpuid.CPU.Has(cpuid.SHA2) {
 		return true
@@ -42,5 +47,4 @@ func hasArmSha2() bool {
 		return false
 	}
 	return bytes.Contains(cpuInfo, []byte(sha256Feature))
-
 }

go.mod

@@ -2,4 +2,4 @@ module github.com/minio/sha256-simd
 
 go 1.13
 
-require github.com/klauspost/cpuid/v2 v2.0.6
+require github.com/klauspost/cpuid/v2 v2.2.3

go.sum

@@ -1,2 +1,4 @@
-github.com/klauspost/cpuid/v2 v2.0.6 h1:dQ5ueTiftKxp0gyjKSx5+8BtPWkyQbd95m8Gys/RarI=
-github.com/klauspost/cpuid/v2 v2.0.6/go.mod h1:FInQzS24/EEf25PyTYn52gqo7WaD8xa0213Md/qVLRg=
+github.com/klauspost/cpuid/v2 v2.2.3 h1:sxCkb+qR91z4vsqw4vGGZlDgPz3G7gjaLyK3V8y70BU=
+github.com/klauspost/cpuid/v2 v2.2.3/go.mod h1:RVVoqg1df56z8g3pUjL/3lE5UfnlrJX8tyFgg4nqhuY=
+golang.org/x/sys v0.0.0-20220704084225-05e143d24a9e h1:CsOuNlbOuf0mzxJIefr6Q4uAUetRUwZE4qt7VfzP+xo=
+golang.org/x/sys v0.0.0-20220704084225-05e143d24a9e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

sha256.go

@@ -20,9 +20,6 @@ import (
 	"crypto/sha256"
 	"encoding/binary"
 	"hash"
-	"runtime"
-
-	"github.com/klauspost/cpuid/v2"
 )
 
 // Size - The size of a SHA256 checksum in bytes.
@@ -68,42 +65,34 @@ func (d *digest) Reset() {
 type blockfuncType int
 
 const (
-	blockfuncGeneric blockfuncType = iota
-	blockfuncSha     blockfuncType = iota
-	blockfuncArm     blockfuncType = iota
+	blockfuncStdlib blockfuncType = iota
+	blockfuncIntelSha
+	blockfuncArmSha2
+	blockfuncForceGeneric = -1
 )
 
 var blockfunc blockfuncType
 
 func init() {
-	blockfunc = blockfuncGeneric
 	switch {
-	case hasSHAExtensions():
-		blockfunc = blockfuncSha
+	case hasIntelSha:
+		blockfunc = blockfuncIntelSha
 	case hasArmSha2():
-		blockfunc = blockfuncArm
-	default:
-		blockfunc = blockfuncGeneric
+		blockfunc = blockfuncArmSha2
 	}
 }
 
-var avx512 = cpuid.CPU.Supports(cpuid.AVX512F, cpuid.AVX512DQ, cpuid.AVX512BW, cpuid.AVX512VL)
-
-// hasSHAExtensions return  whether the cpu supports SHA extensions.
-func hasSHAExtensions() bool {
-	return cpuid.CPU.Supports(cpuid.SHA, cpuid.SSSE3, cpuid.SSE4) && runtime.GOARCH == "amd64"
-}
-
 // New returns a new hash.Hash computing the SHA256 checksum.
 func New() hash.Hash {
-	if blockfunc != blockfuncGeneric {
-		d := new(digest)
-		d.Reset()
-		return d
+	if blockfunc == blockfuncStdlib {
+		// Fallback to the standard golang implementation
+		// if no features were found.
+		return sha256.New()
 	}
-	// Fallback to the standard golang implementation
-	// if no features were found.
-	return sha256.New()
+
+	d := new(digest)
+	d.Reset()
+	return d
 }
 
 // Sum256 - single caller sha256 helper
@@ -272,11 +261,11 @@ func (d *digest) checkSum() (digest [Size]byte) {
 }
 
 func block(dig *digest, p []byte) {
-	if blockfunc == blockfuncSha {
-		blockShaGo(dig, p)
-	} else if blockfunc == blockfuncArm {
-		blockArmGo(dig, p)
-	} else if blockfunc == blockfuncGeneric {
+	if blockfunc == blockfuncIntelSha {
+		blockIntelShaGo(dig, p)
+	} else if blockfunc == blockfuncArmSha2 {
+		blockArmSha2Go(dig, p)
+	} else {
 		blockGeneric(dig, p)
 	}
 }

sha256_test.go

@@ -54,8 +54,6 @@ import (
 	"fmt"
 	"strings"
 	"testing"
-
-	"github.com/klauspost/cpuid/v2"
 )
 
 type sha256Test struct {
@@ -2217,7 +2215,7 @@ func TestGolden(t *testing.T) {
 	}()
 
 	if true {
-		blockfunc = blockfuncGeneric
+		blockfunc = blockfuncForceGeneric
 		for _, g := range golden {
 			s := fmt.Sprintf("%x", Sum256([]byte(g.in)))
 			if Sum256([]byte(g.in)) != g.out {
@@ -2226,8 +2224,8 @@ func TestGolden(t *testing.T) {
 		}
 	}
 
-	if cpuid.CPU.Supports(cpuid.SHA, cpuid.SSSE3, cpuid.SSE4) {
-		blockfunc = blockfuncSha
+	if hasIntelSha {
+		blockfunc = blockfuncIntelSha
 		for _, g := range golden {
 			s := fmt.Sprintf("%x", Sum256([]byte(g.in)))
 			if Sum256([]byte(g.in)) != g.out {
@@ -2237,7 +2235,7 @@ func TestGolden(t *testing.T) {
 	}
 
 	if hasArmSha2() {
-		blockfunc = blockfuncArm
+		blockfunc = blockfuncArmSha2
 		for _, g := range golden {
 			s := fmt.Sprintf("%x", Sum256([]byte(g.in)))
 			if Sum256([]byte(g.in)) != g.out {
@@ -2269,27 +2267,34 @@ func benchmarkSize(b *testing.B, size int) {
 	b.ResetTimer()
 	for i := 0; i < b.N; i++ {
 		bench.Reset()
-		bench.Write(buf[:size])
+		bench.Write(buf)
 		bench.Sum(sum[:0])
 	}
 }
 
 func BenchmarkHash(b *testing.B) {
-	algos := []struct {
+	type alg struct {
 		n string
 		t blockfuncType
-		f bool
-	}{
-		{"SHA_", blockfuncSha, hasSHAExtensions()},
-		{"GEN_", blockfuncGeneric, true},
 	}
+	algos := make([]alg, 0, 2)
+
+	algos = append(algos, alg{"Generic", blockfuncForceGeneric})
+	if hasIntelSha {
+		algos = append(algos, alg{"IntelSHA", blockfuncIntelSha})
+	}
+	if hasArmSha2() {
+		algos = append(algos, alg{"ArmSha2", blockfuncArmSha2})
+	}
+	algos = append(algos, alg{"GoStdlib", blockfuncStdlib})
 
 	sizes := []struct {
 		n string
 		f func(*testing.B, int)
 		s int
 	}{
 		{"8Bytes", benchmarkSize, 1 << 3},
+		{"64Bytes", benchmarkSize, 1 << 6},
 		{"1K", benchmarkSize, 1 << 10},
 		{"8K", benchmarkSize, 1 << 13},
 		{"1M", benchmarkSize, 1 << 20},
@@ -2298,14 +2303,15 @@ func BenchmarkHash(b *testing.B) {
 	}
 
 	for _, a := range algos {
-		if a.f {
-			blockfuncSaved := blockfunc
+		func() {
+			orig := blockfunc
+			defer func() { blockfunc = orig }()
+
 			blockfunc = a.t
 			for _, y := range sizes {
 				s := a.n + "/" + y.n
 				b.Run(s, func(b *testing.B) { y.f(b, y.s) })
 			}
-			blockfunc = blockfuncSaved
-		}
+		}()
 	}
 }

sha256blockAvx512_amd64.go

@@ -1,4 +1,5 @@
-//+build !noasm,!appengine,gc
+//go:build !noasm && !appengine && gc
+// +build !noasm,!appengine,gc
 
 /*
  * Minio Cloud Storage, (C) 2017 Minio, Inc.

sha256blockAvx512_amd64_test.go

@@ -1,4 +1,5 @@
-//+build !noasm,!appengine,gc
+//go:build !noasm && !appengine && gc
+// +build !noasm,!appengine,gc
 
 /*
  * Minio Cloud Storage, (C) 2017 Minio, Inc.
@@ -31,7 +32,7 @@ import (
 
 func TestGoldenAVX512(t *testing.T) {
 
-	if !avx512 {
+	if !hasAvx512 {
 		t.SkipNow()
 		return
 	}
@@ -75,7 +76,7 @@ func initDigests() *[512]byte {
 
 func testSha256Avx512(t *testing.T, offset, padding int) [16][]byte {
 
-	if !avx512 {
+	if !hasAvx512 {
 		t.SkipNow()
 		return [16][]byte{}
 	}
@@ -118,7 +119,7 @@ func TestAvx512_3Blocks(t *testing.T) { testSha256Avx512(t, 47, 55) }
 
 func TestAvx512_MixedBlocks(t *testing.T) {
 
-	if !avx512 {
+	if !hasAvx512 {
 		t.SkipNow()
 		return
 	}
@@ -153,7 +154,7 @@ func TestAvx512_MixedBlocks(t *testing.T) {
 
 func TestAvx512_MixedWithNilBlocks(t *testing.T) {
 
-	if !avx512 {
+	if !hasAvx512 {
 		t.SkipNow()
 		return
 	}
@@ -200,7 +201,7 @@ func TestAvx512_MixedWithNilBlocks(t *testing.T) {
 
 func TestAvx512Server(t *testing.T) {
 
-	if !avx512 {
+	if !hasAvx512 {
 		t.SkipNow()
 		return
 	}
@@ -251,7 +252,7 @@ func TestAvx512Server(t *testing.T) {
 
 func TestAvx512Digest(t *testing.T) {
 
-	if !avx512 {
+	if !hasAvx512 {
 		t.SkipNow()
 		return
 	}
@@ -295,7 +296,7 @@ func benchmarkAvx512SingleCore(h512 []hash.Hash, body []byte) {
 
 func benchmarkAvx512(b *testing.B, size int) {
 
-	if !avx512 {
+	if !hasAvx512 {
 		b.SkipNow()
 		return
 	}
@@ -325,7 +326,7 @@ func BenchmarkAvx512_10M(b *testing.B) { benchmarkAvx512(b, 10*1024*1024) }
 
 func benchmarkAvx512MultiCore(b *testing.B, size, cores int) {
 
-	if !avx512 {
+	if !hasAvx512 {
 		b.SkipNow()
 		return
 	}

sha256block_amd64.go

@@ -1,4 +1,5 @@
-//+build !noasm,!appengine,gc
+//go:build !noasm && !appengine && gc
+// +build !noasm,!appengine,gc
 
 /*
  * Minio Cloud Storage, (C) 2016 Minio, Inc.
@@ -18,10 +19,13 @@
 
 package sha256
 
-func blockArmGo(dig *digest, p []byte) {
-	panic("blockArmGo called unexpectedly")
+func blockArmSha2Go(dig *digest, p []byte) {
+	panic("blockArmSha2Go called unexpectedly")
 }
 
-func blockShaGo(dig *digest, p []byte) {
-	blockSha(&dig.h, p)
+//go:noescape
+func blockIntelSha(h *[8]uint32, message []uint8)
+
+func blockIntelShaGo(dig *digest, p []byte) {
+	blockIntelSha(&dig.h, p)
 }

sha256blockSha_amd64.s → sha256block_amd64.s

@@ -106,7 +106,7 @@ GLOBL SHUF_MASK<>(SB), RODATA|NOPTR, $16
 // X13 saved hash state // CDGH
 // X15 data shuffle mask (constant)
 
-TEXT ·blockSha(SB), NOSPLIT, $0-32
+TEXT ·blockIntelSha(SB), NOSPLIT, $0-32
 	MOVQ      h+0(FP), DX
 	MOVQ      message_base+8(FP), SI
 	MOVQ      message_len+16(FP), DI

sha256blockSha_amd64_test.go → sha256block_amd64_test.go

@@ -1,4 +1,5 @@
-//+build !noasm,!appengine,gc
+//go:build !noasm && !appengine && gc
+// +build !noasm,!appengine,gc
 
 package sha256
 
@@ -20,7 +21,7 @@ func sha256hash(m []byte) (r [32]byte) {
 	h[6] = 0x1f83d9ab
 	h[7] = 0x5be0cd19
 
-	blockSha(&h, m)
+	blockIntelSha(&h, m)
 	l0 := len(m)
 	l := l0 & (BlockSize - 1)
 	m = m[l0-l:]
@@ -31,7 +32,7 @@ func sha256hash(m []byte) (r [32]byte) {
 	k[l] = 0x80
 
 	if l >= 56 {
-		blockSha(&h, k[:])
+		blockIntelSha(&h, k[:])
 		binary.LittleEndian.PutUint64(k[0:8], 0)
 		binary.LittleEndian.PutUint64(k[8:16], 0)
 		binary.LittleEndian.PutUint64(k[16:24], 0)
@@ -41,7 +42,7 @@ func sha256hash(m []byte) (r [32]byte) {
 		binary.LittleEndian.PutUint64(k[48:56], 0)
 	}
 	binary.BigEndian.PutUint64(k[56:64], uint64(l0)<<3)
-	blockSha(&h, k[:])
+	blockIntelSha(&h, k[:])
 
 	binary.BigEndian.PutUint32(r[0:4], h[0])
 	binary.BigEndian.PutUint32(r[4:8], h[1])
@@ -71,7 +72,7 @@ func TestSha0(t *testing.T) {
 }
 
 func TestSha1(t *testing.T) {
-	if hasSHAExtensions() && !runTestSha(sha256hash) {
+	if hasIntelSha && !runTestSha(sha256hash) {
 		t.Errorf("FAILED")
 	}
 }