diff --git a/sm4/aesni_macros_ppc64x.s b/sm4/aesni_macros_ppc64x.s
index 9d6c64f..ac0809d 100644
--- a/sm4/aesni_macros_ppc64x.s
+++ b/sm4/aesni_macros_ppc64x.s
@@ -55,14 +55,14 @@
 // t2 = t3.S2, t2.S2, t1.S2, t0.S2
 // t3 = t3.S3, t2.S3, t1.S3, t0.S3
 #define PRE_TRANSPOSE_MATRIX(T0, T1, T2, T3) \
-	VPERM T0, T1, M0, TMP0; \
-	VPERM T2, T3, M0, TMP1; \
-	VPERM T0, T1, M1, TMP2; \
-	VPERM T2, T3, M1, TMP3; \
-	VPERM TMP0, TMP1, M2, T0; \
-	VPERM TMP0, TMP1, M3, T1; \
-	VPERM TMP2, TMP3, M2, T2; \
-	VPERM TMP2, TMP3, M3, T3
+	VMRGEW T0, T1, TMP0; \
+	VMRGEW T2, T3, TMP1; \
+	VMRGOW T0, T1, TMP2; \
+	VMRGOW T2, T3, TMP3; \
+	XXPERMDI TMP0, TMP1, $0, T0; \
+	XXPERMDI TMP0, TMP1, $3, T2; \
+	XXPERMDI TMP2, TMP3, $0, T1; \
+	XXPERMDI TMP2, TMP3, $3, T3
 
 // input: from high to low
 // t0 = t0.S3, t0.S2, t0.S1, t0.S0
@@ -75,14 +75,14 @@
 // t2 = t0.S2, t1.S2, t2.S2, t3.S2
 // t3 = t0.S3, t1.S3, t2.S3, t3.S3
 #define TRANSPOSE_MATRIX(T0, T1, T2, T3) \
-	VPERM T1, T0, M0, TMP0; \
-	VPERM T1, T0, M1, TMP1; \
-	VPERM T3, T2, M0, TMP2; \
-	VPERM T3, T2, M1, TMP3; \
-	VPERM TMP2, TMP0, M2, T0; \
-	VPERM TMP2, TMP0, M3, T1; \
-	VPERM TMP3, TMP1, M2, T2; \
-	VPERM TMP3, TMP1, M3, T3; \	
+	VMRGEW T1, T0, TMP0; \
+	VMRGEW T3, T2, TMP1; \
+	VMRGOW T1, T0, TMP2; \
+	VMRGOW T3, T2, TMP3; \
+	XXPERMDI TMP1, TMP0, $0, T0; \
+	XXPERMDI TMP1, TMP0, $3, T2; \
+	XXPERMDI TMP3, TMP2, $0, T1; \
+	XXPERMDI TMP3, TMP2, $3, T3
 
 // Affine Transform
 // parameters: