diff --git a/chacha20poly1305/chacha20poly1305.go b/chacha20poly1305/chacha20poly1305.go
index 3f0dcb9..e28f49d 100644
--- a/chacha20poly1305/chacha20poly1305.go
+++ b/chacha20poly1305/chacha20poly1305.go
@@ -7,6 +7,7 @@ package chacha20poly1305 // import "golang.org/x/crypto/chacha20poly1305"
 
 import (
 	"crypto/cipher"
+	"encoding/binary"
 	"errors"
 )
 
@@ -18,7 +19,7 @@ const (
 )
 
 type chacha20poly1305 struct {
-	key [32]byte
+	key [8]uint32
 }
 
 // New returns a ChaCha20-Poly1305 AEAD that uses the given, 256-bit key.
@@ -27,7 +28,14 @@ func New(key []byte) (cipher.AEAD, error) {
 		return nil, errors.New("chacha20poly1305: bad key length")
 	}
 	ret := new(chacha20poly1305)
-	copy(ret.key[:], key)
+	ret.key[0] = binary.LittleEndian.Uint32(key[0:4])
+	ret.key[1] = binary.LittleEndian.Uint32(key[4:8])
+	ret.key[2] = binary.LittleEndian.Uint32(key[8:12])
+	ret.key[3] = binary.LittleEndian.Uint32(key[12:16])
+	ret.key[4] = binary.LittleEndian.Uint32(key[16:20])
+	ret.key[5] = binary.LittleEndian.Uint32(key[20:24])
+	ret.key[6] = binary.LittleEndian.Uint32(key[24:28])
+	ret.key[7] = binary.LittleEndian.Uint32(key[28:32])
 	return ret, nil
 }
 
diff --git a/chacha20poly1305/chacha20poly1305_amd64.go b/chacha20poly1305/chacha20poly1305_amd64.go
index 7cd7ad8..17f94e9 100644
--- a/chacha20poly1305/chacha20poly1305_amd64.go
+++ b/chacha20poly1305/chacha20poly1305_amd64.go
@@ -72,20 +72,20 @@ func isSet(bitpos uint, value uint32) bool {
 
 // setupState writes a ChaCha20 input matrix to state. See
 // https://tools.ietf.org/html/rfc7539#section-2.3.
-func setupState(state *[16]uint32, key *[32]byte, nonce []byte) {
+func setupState(state *[16]uint32, key *[8]uint32, nonce []byte) {
 	state[0] = 0x61707865
 	state[1] = 0x3320646e
 	state[2] = 0x79622d32
 	state[3] = 0x6b206574
 
-	state[4] = binary.LittleEndian.Uint32(key[:4])
-	state[5] = binary.LittleEndian.Uint32(key[4:8])
-	state[6] = binary.LittleEndian.Uint32(key[8:12])
-	state[7] = binary.LittleEndian.Uint32(key[12:16])
-	state[8] = binary.LittleEndian.Uint32(key[16:20])
-	state[9] = binary.LittleEndian.Uint32(key[20:24])
-	state[10] = binary.LittleEndian.Uint32(key[24:28])
-	state[11] = binary.LittleEndian.Uint32(key[28:32])
+	state[4] = key[0]
+	state[5] = key[1]
+	state[6] = key[2]
+	state[7] = key[3]
+	state[8] = key[4]
+	state[9] = key[5]
+	state[10] = key[6]
+	state[11] = key[7]
 
 	state[12] = 0
 	state[13] = binary.LittleEndian.Uint32(nonce[:4])
diff --git a/chacha20poly1305/chacha20poly1305_generic.go b/chacha20poly1305/chacha20poly1305_generic.go
index 4ac014f..8d28ce2 100644
--- a/chacha20poly1305/chacha20poly1305_generic.go
+++ b/chacha20poly1305/chacha20poly1305_generic.go
@@ -16,15 +16,17 @@ func roundTo16(n int) int {
 }
 
 func (c *chacha20poly1305) sealGeneric(dst, nonce, plaintext, additionalData []byte) []byte {
-	var counter [16]byte
-	copy(counter[4:], nonce)
+	ret, out := sliceForAppend(dst, len(plaintext)+poly1305.TagSize)
 
 	var polyKey [32]byte
-	chacha20.XORKeyStream(polyKey[:], polyKey[:], &counter, &c.key)
-
-	ret, out := sliceForAppend(dst, len(plaintext)+poly1305.TagSize)
-	counter[0] = 1
-	chacha20.XORKeyStream(out, plaintext, &counter, &c.key)
+	s := chacha20.New(c.key, [3]uint32{
+		binary.LittleEndian.Uint32(nonce[0:4]),
+		binary.LittleEndian.Uint32(nonce[4:8]),
+		binary.LittleEndian.Uint32(nonce[8:12]),
+	})
+	s.XORKeyStream(polyKey[:], polyKey[:])
+	s.Advance() // skip the next 32 bytes
+	s.XORKeyStream(out, plaintext)
 
 	polyInput := make([]byte, roundTo16(len(additionalData))+roundTo16(len(plaintext))+8+8)
 	copy(polyInput, additionalData)
@@ -44,11 +46,14 @@ func (c *chacha20poly1305) openGeneric(dst, nonce, ciphertext, additionalData []
 	copy(tag[:], ciphertext[len(ciphertext)-16:])
 	ciphertext = ciphertext[:len(ciphertext)-16]
 
-	var counter [16]byte
-	copy(counter[4:], nonce)
-
 	var polyKey [32]byte
-	chacha20.XORKeyStream(polyKey[:], polyKey[:], &counter, &c.key)
+	s := chacha20.New(c.key, [3]uint32{
+		binary.LittleEndian.Uint32(nonce[0:4]),
+		binary.LittleEndian.Uint32(nonce[4:8]),
+		binary.LittleEndian.Uint32(nonce[8:12]),
+	})
+	s.XORKeyStream(polyKey[:], polyKey[:])
+	s.Advance() // skip the next 32 bytes
 
 	polyInput := make([]byte, roundTo16(len(additionalData))+roundTo16(len(ciphertext))+8+8)
 	copy(polyInput, additionalData)
@@ -64,7 +69,6 @@ func (c *chacha20poly1305) openGeneric(dst, nonce, ciphertext, additionalData []
 		return nil, errOpen
 	}
 
-	counter[0] = 1
-	chacha20.XORKeyStream(out, ciphertext, &counter, &c.key)
+	s.XORKeyStream(out, ciphertext)
 	return ret, nil
 }