diff --git a/sm3/sm3block_amd64.s b/sm3/sm3block_amd64.s
index ea07cf3..6d082ee 100644
--- a/sm3/sm3block_amd64.s
+++ b/sm3/sm3block_amd64.s
@@ -1052,10 +1052,10 @@ avx2_loop0: // at each iteration works with one block (512 bit)
 	VPSHUFB BYTE_FLIP_MASK, XTMP3, XTMP3
 
 	// Transpose data into high/low parts
-	VPERM2I128 $0x20, XTMP2, XTMP0, XDWORD0 // w3, w2, w1, w0
-	VPERM2I128 $0x31, XTMP2, XTMP0, XDWORD1 // w7, w6, w5, w4
-	VPERM2I128 $0x20, XTMP3, XTMP1, XDWORD2 // w11, w10, w9, w8
-	VPERM2I128 $0x31, XTMP3, XTMP1, XDWORD3 // w15, w14, w13, w12
+	VPERM2I128 $0x20, XTMP2, XTMP0, XDWORD0 // w19, w18, w17, w16;  w3,  w2,  w1,  w0
+	VPERM2I128 $0x31, XTMP2, XTMP0, XDWORD1 // w23, w22, w21, w20;  w7,  w6,  w5,  w4
+	VPERM2I128 $0x20, XTMP3, XTMP1, XDWORD2 // w27, w26, w25, w24; w11, w10,  w9,  w8
+	VPERM2I128 $0x31, XTMP3, XTMP1, XDWORD3 // w31, w30, w29, w28; w15, w14, w13, w12
 
 avx2_last_block_enter:
 	ADDQ $64, INP
diff --git a/sm4/asm_amd64.s b/sm4/asm_amd64.s
index fb58739..005f4c4 100644
--- a/sm4/asm_amd64.s
+++ b/sm4/asm_amd64.s
@@ -113,11 +113,11 @@ GLOBL fk_mask<>(SB), RODATA, $16
   MOVOU x, XTMP6;                     \
   PSLLL $13, XTMP6;                   \
   PSRLL $19, y;                       \
-  PXOR XTMP6, y;                      \ //y = X roll 13  
+  POR XTMP6, y;                      \ //y = X roll 13  
   PSLLL $10, XTMP6;                   \
   MOVOU x, XTMP7;                     \
   PSRLL $9, XTMP7;                    \
-  PXOR XTMP6, XTMP7;                  \ //XTMP7 = x roll 23
+  POR XTMP6, XTMP7;                  \ //XTMP7 = x roll 23
   PXOR XTMP7, y;                      \
   PXOR y, x                        
 
diff --git a/sm4/cipher_asm_fuzzy_test.go b/sm4/cipher_asm_fuzzy_test.go
new file mode 100644
index 0000000..661a686
--- /dev/null
+++ b/sm4/cipher_asm_fuzzy_test.go
@@ -0,0 +1,45 @@
+//go:build amd64
+// +build amd64
+
+package sm4
+
+import (
+	"crypto/rand"
+	"io"
+	"reflect"
+	"testing"
+	"time"
+)
+
+func TestExpandKey(t *testing.T) {
+	key := make([]byte, 16)
+
+	encRes1 := make([]uint32, 32)
+	decRes1 := make([]uint32, 32)
+	encRes2 := make([]uint32, 32)
+	decRes2 := make([]uint32, 32)
+	var timeout *time.Timer
+
+	if testing.Short() {
+		timeout = time.NewTimer(10 * time.Millisecond)
+	} else {
+		timeout = time.NewTimer(2 * time.Second)
+	}
+
+	for {
+		select {
+		case <-timeout.C:
+			return
+		default:
+		}
+		io.ReadFull(rand.Reader, key)
+		expandKeyGo(key, encRes1, decRes1)
+		expandKeyAsm(&key[0], &ck[0], &encRes2[0], &decRes2[0])
+		if !reflect.DeepEqual(encRes1, encRes2) {
+			t.Fatalf("expected=%v, result=%v\n", encRes1, encRes2)
+		}
+		if !reflect.DeepEqual(decRes1, decRes2) {
+			t.Fatalf("expected=%v, result=%v\n", encRes1, encRes2)
+		}
+	}
+}