diff --git a/sm3/sm3blocks_s390x.s b/sm3/sm3blocks_s390x.s
index 358ecf9..b716f6f 100644
--- a/sm3/sm3blocks_s390x.s
+++ b/sm3/sm3blocks_s390x.s
@@ -44,17 +44,23 @@ TEXT ·transposeMatrix(SB),NOSPLIT,$0
 	VL (R2), V6
 	VL 16(R2), V7
 
+	MOVD $mask<>+0x00(SB), R2
+	VLM (R2), V8, V11
+
+	TRANSPOSE_MATRIX(V0, V2, V4, V6, V8, V9, V10, V11, V12, V13, V14, V15)
+	TRANSPOSE_MATRIX(V1, V3, V5, V7, V8, V9, V10, V11, V12, V13, V14, V15)
+
 	MOVD 	0(R1), R2
-	VST V6, (R2)
-	VST V7, 16(R2)
-	MOVD 	8(R1), R2
-	VST V4, (R2)
-	VST V5, 16(R2)
-	MOVD 	16(R1), R2
-	VST V2, (R2)
-	VST V3, 16(R2)
-	MOVD 	24(R1), R2
 	VST V0, (R2)
 	VST V1, 16(R2)
+	MOVD 	8(R1), R2
+	VST V2, (R2)
+	VST V3, 16(R2)
+	MOVD 	16(R1), R2
+	VST V4, (R2)
+	VST V5, 16(R2)
+	MOVD 	24(R1), R2
+	VST V6, (R2)
+	VST V7, 16(R2)
 
 	RET
diff --git a/sm3/sm3blocks_s390x_test.go b/sm3/sm3blocks_s390x_test.go
index f785466..94021ef 100644
--- a/sm3/sm3blocks_s390x_test.go
+++ b/sm3/sm3blocks_s390x_test.go
@@ -24,7 +24,6 @@ func TestTransposeMatrix(t *testing.T) {
 	fmt.Println()
 	for i := 0; i < 4; i++ {
 		for j := 0; j < 8; j++ {
-			m[i][j] = uint32(i*4 + j)
 			fmt.Printf("%04x ", m[i][j])
 		}
 		fmt.Println()