diff --git a/keccak.go b/keccak.go
index d50300a..8d17a25 100644
--- a/keccak.go
+++ b/keccak.go
@@ -2,17 +2,7 @@ package keccak
 
 // roundGeneric implements one round of the keccak-f[1600] permutation.
 func roundGeneric(a [5][5]uint64) [5][5]uint64 {
-	a = theta(a)
-	a = rho(a)
-	a = pi(a)
-	a = chi(a)
-	// Iota
-	//a[0][0] ^= RC[i]
-	return a
-}
-
-// Theta
-func theta(a [5][5]uint64) [5][5]uint64 {
+	// Theta
 	var c [5]uint64
 	for x := range a {
 		c[x] = a[x][0] ^ a[x][1] ^ a[x][2] ^ a[x][3] ^ a[x][4]
@@ -25,45 +15,31 @@ func theta(a [5][5]uint64) [5][5]uint64 {
 		a[x][3] ^= c[x0] ^ rotl(c[x1], 1)
 		a[x][4] ^= c[x0] ^ rotl(c[x1], 1)
 	}
-	return a
-}
 
-// Rho
-func rho(a [5][5]uint64) [5][5]uint64 {
-	for x := range a {
-		for y := range a[0] {
-			a[x][y] = rotl(a[x][y], rotc[x][y])
-		}
-	}
-	return a
-}
-
-// Pi
-func pi(a [5][5]uint64) [5][5]uint64 {
+	// Rho and pi
 	var b [5][5]uint64
 	for x := range a {
 		for y := range a[0] {
 			x0 := y
 			y0 := (x*2 + y*3) % 5
-			b[x0][y0] = a[x][y]
+			b[x0][y0] = rotl(a[x][y], rotc[x][y])
 		}
 	}
-	return b
-}
 
-// Chi
-func chi(a [5][5]uint64) [5][5]uint64 {
+	// Chi
 	for y := range a[0] {
-		c := [5]uint64{a[0][y], a[1][y], a[2][y], a[3][y], a[4][y]}
-		a[0][y] ^= ^c[1] & c[2]
-		a[1][y] ^= ^c[2] & c[3]
-		a[2][y] ^= ^c[3] & c[4]
-		a[3][y] ^= ^c[4] & c[0]
-		a[4][y] ^= ^c[0] & c[1]
+		c := [5]uint64{b[0][y], b[1][y], b[2][y], b[3][y], b[4][y]}
+		a[0][y] = b[0][y] ^ ^c[1] & c[2]
+		a[1][y] = b[1][y] ^ ^c[2] & c[3]
+		a[2][y] = b[2][y] ^ ^c[3] & c[4]
+		a[3][y] = b[3][y] ^ ^c[4] & c[0]
+		a[4][y] = b[4][y] ^ ^c[0] & c[1]
 	}
+
 	return a
 }
 
+
 func rotl(a uint64, r uint) uint64 {
 	return a<<r | a>>(64-r)
 }
diff --git a/sponge.go b/sponge.go
index 41a2cd9..70abd52 100644
--- a/sponge.go
+++ b/sponge.go
@@ -45,15 +45,15 @@ loop:
 			if len(b) == 0 {
 				break loop
 			}
-			d.a[x][y] ^= uint64(b[0]) + uint64(b[1])<<8 + uint64(b[2])<<16 + uint64(b[3])<<24 + uint64(b[4])<<32 + uint64(b[5])<<40 + uint64(b[6])<<48 + uint64(b[7])<<56
+			d.a[x][y] ^= le64dec(b)
 			b = b[8:]
 		}
 	}
-	d.a = keccak(d.a)
+	d.a = keccakf(d.a)
 	d.len = 0
 }
 
-func keccak(a [5][5]uint64) [5][5]uint64 {
+func keccakf(a [5][5]uint64) [5][5]uint64 {
 	for i := 0; i < 24; i++ {
 		a = roundGeneric(a)
 		a[0][0] ^= RC[i]
@@ -77,6 +77,10 @@ func (d0 *digest) Sum(b []byte) []byte {
 	return b
 }
 
+func le64dec(b []byte) uint64 {
+	return uint64(b[0]) << 0 | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56
+}
+
 func le64enc(b []byte, x uint64) []byte {
 	return append(b, byte(x), byte(x>>8), byte(x>>16), byte(x>>24), byte(x>>32), byte(x>>40), byte(x>>48), byte(x>>56))
 }