diff --git a/hash/shake/hashes_generic.go b/hash/shake/hashes_generic.go
deleted file mode 100644
index c6da68a..0000000
--- a/hash/shake/hashes_generic.go
+++ /dev/null
@@ -1,27 +0,0 @@
-// Copyright 2017 The Go Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-//+build gccgo appengine !s390x
-
-package shake
-
-import (
-	"hash"
-)
-
-// new224Asm returns an assembly implementation of SHA3-224 if available,
-// otherwise it returns nil.
-func new224Asm() hash.Hash { return nil }
-
-// new256Asm returns an assembly implementation of SHA3-256 if available,
-// otherwise it returns nil.
-func new256Asm() hash.Hash { return nil }
-
-// new384Asm returns an assembly implementation of SHA3-384 if available,
-// otherwise it returns nil.
-func new384Asm() hash.Hash { return nil }
-
-// new512Asm returns an assembly implementation of SHA3-512 if available,
-// otherwise it returns nil.
-func new512Asm() hash.Hash { return nil }
diff --git a/hash/shake/sha3.go b/hash/shake/sha3.go
index 2c01bb8..8c8a945 100644
--- a/hash/shake/sha3.go
+++ b/hash/shake/sha3.go
@@ -83,7 +83,7 @@ func (d *state) permute() {
 	case spongeAbsorbing:
 		// If we're absorbing, we need to xor the input into the state
 		// before applying the permutation.
-		xorIn(d, d.buf)
+		xorInUnaligned(d, d.buf)
 		d.buf = d.storage[:0]
 		keccakF1600(&d.a)
 	case spongeSqueezing:
@@ -91,7 +91,7 @@ func (d *state) permute() {
 		// copying more output.
 		keccakF1600(&d.a)
 		d.buf = d.storage[:d.rate]
-		copyOut(d, d.buf)
+		copyOutUnaligned(d, d.buf)
 	}
 }
 
@@ -119,7 +119,7 @@ func (d *state) padAndPermute(dsbyte byte) {
 	d.permute()
 	d.state = spongeSqueezing
 	d.buf = d.storage[:d.rate]
-	copyOut(d, d.buf)
+	copyOutUnaligned(d, d.buf)
 }
 
 // Write absorbs more data into the hash's state. It produces an error
@@ -136,7 +136,7 @@ func (d *state) Write(p []byte) (written int, err error) {
 	for len(p) > 0 {
 		if len(d.buf) == 0 && len(p) >= d.rate {
 			// The fast path; absorb a full "rate" bytes of input and apply the permutation.
-			xorIn(d, p[:d.rate])
+			xorInUnaligned(d, p[:d.rate])
 			p = p[d.rate:]
 			keccakF1600(&d.a)
 		} else {
diff --git a/hash/shake/sha3_test.go b/hash/shake/sha3_test.go
index 9176a20..1480837 100644
--- a/hash/shake/sha3_test.go
+++ b/hash/shake/sha3_test.go
@@ -22,6 +22,11 @@ import (
 	"testing"
 )
 
+var (
+	xorIn   = xorInUnaligned
+	copyOut = copyOutUnaligned
+)
+
 const (
 	testString  = "brekeccakkeccak koax koax"
 	katFilename = "testdata/keccakKats.json.deflate"
@@ -68,7 +73,7 @@ func testUnalignedAndGeneric(t *testing.T, testf func(impl string)) {
 	xorIn, copyOut = xorInGeneric, copyOutGeneric
 	testf("generic")
 	if xorImplementationUnaligned != "generic" {
-		xorIn, copyOut = xorInUnaligned, copyOutUnaligned
+		xorIn, copyOut = xorInGeneric, copyOutGeneric
 		testf("unaligned")
 	}
 	xorIn, copyOut = xorInOrig, copyOutOrig
@@ -336,7 +341,7 @@ func benchmarkShake(b *testing.B, h *CShake, size, num int) {
 	b.StopTimer()
 	h.Reset()
 	data := sequentialBytes(size)
-	d := make([]byte, 32)
+	var d [32]byte
 
 	b.SetBytes(int64(size * num))
 	b.StartTimer()
@@ -346,7 +351,7 @@ func benchmarkShake(b *testing.B, h *CShake, size, num int) {
 		for j := 0; j < num; j++ {
 			h.Write(data)
 		}
-		h.Read(d)
+		h.Read(d[:])
 	}
 }
 
@@ -354,6 +359,53 @@ func BenchmarkShake128_MTU(b *testing.B)  { benchmarkShake(b, NewShake128(), 135
 func BenchmarkShake256_MTU(b *testing.B)  { benchmarkShake(b, NewShake256(), 1350, 1) }
 func BenchmarkShake256_16x(b *testing.B)  { benchmarkShake(b, NewShake256(), 16, 1024) }
 func BenchmarkShake256_1MiB(b *testing.B) { benchmarkShake(b, NewShake256(), 1024, 1024) }
+func BenchmarkCShake256_1MiB(b *testing.B) {
+	benchmarkShake(b, NewCShake256([]byte("ABC"), nil), 1024, 1024)
+}
+
+func BenchmarkCShakeSum(b *testing.B) {
+	//benchmarkShake(b, NewCShake256([]byte("ABC"), nil), 1024, 1024)
+	data := sequentialBytes(1024)
+	var d [32]byte
+
+	//var c CShake
+	//c.Init(CSHAKE_256, )
+	c := NewCShake256([]byte("ABC"), nil)
+	for i := 0; i < b.N; i++ {
+		c.CShakeSum(d[:], data)
+	}
+}
+
+func BenchmarkCShakeSum3(b *testing.B) {
+	//benchmarkShake(b, NewCShake256([]byte("ABC"), nil), 1024, 1024)
+	data := sequentialBytes(1024)
+	var d [32]byte
+
+	c := NewCShake256([]byte("ABC"), nil)
+	for i := 0; i < b.N; i++ {
+		c.Reset()
+		for i := 0; i < 10; i++ {
+			c.Write(data)
+		}
+		c.Read(d[:])
+	}
+}
+
+func BenchmarkCShakeSum4(b *testing.B) {
+	//benchmarkShake(b, NewCShake256([]byte("ABC"), nil), 1024, 1024)
+	data := sequentialBytes(1024)
+	var d [32]byte
+
+	var c CShake //:= NewCShake256([]byte("ABC"), nil)
+	c.Init(CSHAKE_256, []byte("ABC"), nil)
+	for i := 0; i < b.N; i++ {
+		c.Reset()
+		for i := 0; i < 10; i++ {
+			c.Write(data)
+		}
+		c.Read(d[:])
+	}
+}
 
 func Example_sum() {
 	buf := []byte("some data to hash")
@@ -414,3 +466,35 @@ func ExampleCShake256() {
 	//a90a4c6ca9af2156eba43dc8398279e6b60dcd56fb21837afe6c308fd4ceb05b9dd98c6ee866ca7dc5a39d53e960f400bcd5a19c8a2d6ec6459f63696543a0d8
 	//85e73a72228d08b46515553ca3a29d47df3047e5d84b12d6c2c63e579f4fd1105716b7838e92e981863907f434bfd4443c9e56ea09da998d2f9b47db71988109
 }
+
+func TestDoubleInit(t *testing.T) {
+	var h1 CShake
+
+	h1.Init(CSHAKE_256, []byte("ABC"), []byte("DEF"))
+	h2 := NewCShake256([]byte("ABC"), []byte("DEF"))
+
+	h1.Write([]byte("FUCKOFF"))
+	h2.Write([]byte("FUCKOFF"))
+
+	var out1, out2 [32]byte
+	h1.Read(out1[:])
+	h2.Read(out2[:])
+
+	if !bytes.Equal(out1[:], out2[:]) {
+		t.FailNow()
+	}
+}
+
+func BenchmarkCShakeSum2(b *testing.B) {
+	var f = []byte{0x02, 0x00}
+	var o1 [32]byte
+	var i1 [449]byte
+
+	var c CShake
+	c.Init(CSHAKE_256, nil, f)
+
+	for i := 0; i < b.N; i++ {
+		c.CShakeSum(o1[:], i1[:])
+	}
+
+}
diff --git a/hash/shake/shake.go b/hash/shake/shake.go
index 726d97a..de7e83d 100644
--- a/hash/shake/shake.go
+++ b/hash/shake/shake.go
@@ -131,6 +131,65 @@ func NewCShake256(N, S []byte) *CShake {
 	return newCShake(N, S, rate256, dsbyteCShake)
 }
 
+const (
+	CSHAKE_256 = 0
+	CSHAKE_128 = 1
+	SHAKE_128  = 2
+	SHAKE_256  = 2
+)
+
+func (c *CShake) Init(id int, N, S []byte) {
+	// leftEncode returns max 9 bytes
+	initBlock := make([]byte, 0, 9*2+len(N)+len(S))
+	initBlock = append(initBlock, leftEncode(uint64(len(N)*8))...)
+	initBlock = append(initBlock, N...)
+	initBlock = append(initBlock, leftEncode(uint64(len(S)*8))...)
+	initBlock = append(initBlock, S...)
+
+	switch id {
+	case CSHAKE_256:
+		c.state.rate = rate256
+		c.state.dsbyte = dsbyteCShake
+	}
+
+	c.initBlock = bytepad(initBlock, c.state.rate)
+	c.Write(c.initBlock)
+}
+
+func (cs CShake) CShakeSum(out []byte, in []byte) (n int) {
+	d := cs.state
+	blocks := int(len(in) / d.rate)
+
+	// Reset
+	d.buf = d.storage[:0]
+	d.state = spongeAbsorbing
+	for i, _ := range d.a {
+		d.a[i] = 0
+	}
+
+	// Absorb
+	for i := 0; i < blocks; i++ {
+		xorInUnaligned(&d, in[i*d.rate:(i+1)*d.rate])
+		keccakF1600(&d.a)
+	}
+	d.buf = append(d.buf, in[blocks*d.rate:]...)
+	d.padAndPermute(d.dsbyte)
+
+	// Now, do the squeezing.
+	n = len(out)
+	for len(out) > 0 {
+		n := copy(out[:], d.buf)
+		d.buf = d.buf[n:]
+		out = out[n:]
+
+		// Apply the permutation if we've squeezed the sponge dry.
+		if len(d.buf) == 0 {
+			d.permute()
+		}
+	}
+	return n
+}
+
 // ShakeSum128 writes an arbitrary-length digest of data into hash.
 func ShakeSum128(hash, data []byte) {
 	h := NewShake128()
diff --git a/hash/shake/xor.go b/hash/shake/xor.go
deleted file mode 100644
index b6b4370..0000000
--- a/hash/shake/xor.go
+++ /dev/null
@@ -1,16 +0,0 @@
-// Copyright 2015 The Go Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-// +build !amd64,!386,!ppc64le appengine
-
-package shake
-
-var (
-	xorIn            = xorInGeneric
-	copyOut          = copyOutGeneric
-	xorInUnaligned   = xorInGeneric
-	copyOutUnaligned = copyOutGeneric
-)
-
-const xorImplementationUnaligned = "generic"
diff --git a/hash/shake/xor_unaligned.go b/hash/shake/xor_unaligned.go
index d5dd014..972966b 100644
--- a/hash/shake/xor_unaligned.go
+++ b/hash/shake/xor_unaligned.go
@@ -50,9 +50,4 @@ func copyOutUnaligned(d *state, buf []byte) {
 	copy(buf, ab[:])
 }
 
-var (
-	xorIn   = xorInUnaligned
-	copyOut = copyOutUnaligned
-)
-
 const xorImplementationUnaligned = "unaligned"