diff --git a/sm4/gcm_arm64.s b/sm4/gcm_arm64.s
index 5c103bd..adc588e 100644
--- a/sm4/gcm_arm64.s
+++ b/sm4/gcm_arm64.s
@@ -144,8 +144,8 @@ TEXT ·gcmSm4Finish(SB),NOSPLIT,$0
 	VEOR z.B16, x.B16, x.B16;                            \
 	VEOR y.B16, x.B16, x.B16
 
-#define SM4_ROUND(index, RK, x, y, z, z1, z2, t0, t1, t2, t3)  \ 
-	MOVW (index*4)(RK), R19;                          \
+#define SM4_ROUND(RK, x, y, z, z1, z2, t0, t1, t2, t3)  \ 
+	MOVW.P 4(RK), R19;                          \
 	VMOV R19, x.S4;                                   \
 	VEOR t1.B16, x.B16, x.B16;                        \
 	VEOR t2.B16, x.B16, x.B16;                        \
@@ -153,14 +153,14 @@ TEXT ·gcmSm4Finish(SB),NOSPLIT,$0
 	SM4_TAO_L1(x, y, z, z1, z2);                      \
 	VEOR x.B16, t0.B16, t0.B16
 
-// func precomputeTableAsm(productTable *[256]byte, src *[16]byte)
-TEXT ·precomputeTableAsm(SB),NOSPLIT,$0
+// func gcmSm4Init(productTable *[256]byte, rk []uint32)
+TEXT gcmSm4Init(SB),NOSPLIT,$0
 #define pTbl R0
-#define SRC R1
+#define RK R1
 #define I R2
 
 	MOVD productTable+0(FP), pTbl
-	MOVD src+8(FP), SRC
+	MOVD rk+8(FP), RK
 
 	MOVD	$0xC2, I
 	LSL	$56, I
@@ -169,8 +169,27 @@ TEXT ·precomputeTableAsm(SB),NOSPLIT,$0
 	VMOV	I, POLY.D[1]
 	VEOR	ZERO.B16, ZERO.B16, ZERO.B16
 
-	VLD1 (SRC), [B0.B16]
-	VREV64	B0.B16, B0.B16
+	// Encrypt block 0 with the SM4 keys to generate the hash key H
+	VEOR	B0.B16, B0.B16, B0.B16
+	VEOR	B1.B16, B1.B16, B1.B16
+	VEOR	B2.B16, B2.B16, B2.B16
+	VEOR	B3.B16, B3.B16, B3.B16
+	EOR R3, R3
+
+sm4InitEncLoop:	
+	SM4_ROUND(RK, K0, K1, K2, K3, K4, B0, B1, B2, B3)
+	SM4_ROUND(RK, K0, K1, K2, K3, K4, B1, B2, B3, B0)
+	SM4_ROUND(RK, K0, K1, K2, K3, K4, B2, B3, B0, B1)
+	SM4_ROUND(RK, K0, K1, K2, K3, K4, B3, B0, B1, B2)
+
+	ADD $1, R3
+	CMP $8, R3
+	BNE sm4InitEncLoop
+
+	VMOV B0.S[0], B0.S[2]
+	VMOV B1.S[0], B0.S[3]
+	VMOV B2.S[0], B0.S[0]
+	VMOV B3.S[0], B0.S[1]
 
 	// Multiply by 2 modulo P
 	VMOV	B0.D[0], I
@@ -228,7 +247,7 @@ initLoop:
 	BNE	initLoop
 	RET
 #undef I
-#undef SRC
+#undef RK
 #undef pTbl	
 
 // func gcmSm4Data(productTable *[256]byte, data []byte, T *[16]byte)
diff --git a/sm4/sm4_gcm_arm64.go b/sm4/sm4_gcm_arm64.go
index ed4dd8b..9949176 100644
--- a/sm4/sm4_gcm_arm64.go
+++ b/sm4/sm4_gcm_arm64.go
@@ -21,7 +21,7 @@ type sm4CipherGCM struct {
 var _ gcmAble = (*sm4CipherGCM)(nil)
 
 //go:noescape
-func precomputeTableAsm(productTable *[256]byte, src *[16]byte)
+func gcmSm4Init(productTable *[256]byte, rk []uint32)
 
 //go:noescape
 func gcmSm4Data(productTable *[256]byte, data []byte, T *[16]byte)
@@ -41,9 +41,7 @@ func (c *sm4CipherGCM) NewGCM(nonceSize, tagSize int) (cipher.AEAD, error) {
 	g.cipher = &c.sm4CipherAsm
 	g.nonceSize = nonceSize
 	g.tagSize = tagSize
-	var key [gcmBlockSize]byte
-	c.Encrypt(key[:], key[:])
-	precomputeTableAsm(&g.bytesProductTable, &key)
+	gcmSm4Init(&g.bytesProductTable, g.cipher.enc)
 	return g, nil
 }
 
diff --git a/sm4/sm4_gcm_test.go b/sm4/sm4_gcm_test.go
index 4b7c7f1..75acca9 100644
--- a/sm4/sm4_gcm_test.go
+++ b/sm4/sm4_gcm_test.go
@@ -1,5 +1,5 @@
-//go:build amd64
-// +build amd64
+//go:build amd64 || arm64
+// +build amd64 arm64
 
 package sm4