diff --git a/base64_arm64.go b/base64_arm64.go
index d5664d1..dbdc77b 100644
--- a/base64_arm64.go
+++ b/base64_arm64.go
@@ -41,7 +41,7 @@ func encode(enc *Encoding, dst, src []byte) {
 
 func decode(enc *Encoding, dst, src []byte) (int, error) {
 	srcLen := len(src)
-	if srcLen >= 64 {
+	if srcLen >= 24 {
 		remain := srcLen
 		if enc.lut == &encodeStdLut {
 			remain = decodeAsm(dst, src, &dencodeStdLut)
diff --git a/base64_arm64.s b/base64_arm64.s
index a49a652..d15a6ba 100644
--- a/base64_arm64.s
+++ b/base64_arm64.s
@@ -15,6 +15,14 @@ DATA enc_const<>+0x30(SB)/8, $0x1f1e1b1a17161312 // high part of word
 DATA enc_const<>+0x38(SB)/8, $0x0f0e0b0a07060302
 GLOBL enc_const<>(SB), (NOPTR+RODATA), $64
 
+DATA dec_const<>+0x00(SB)/8, $0x0140014001400140 // dec_reshuffle_const0
+DATA dec_const<>+0x08(SB)/8, $0x0140014001400140
+DATA dec_const<>+0x10(SB)/8, $0x0000110000001100 // dec_reshuffle_const1
+DATA dec_const<>+0x18(SB)/8, $0x0000110000001100
+DATA dec_const<>+0x20(SB)/8, $0x090A040506000102 // dec_reshuffle_mask
+DATA dec_const<>+0x28(SB)/8, $0xFFFFFFFF0C0D0E08
+GLOBL dec_const<>(SB), (NOPTR+RODATA), $48
+
 //func encodeAsm(dst, src []byte, lut *[64]byte) int
 TEXT ·encodeAsm(SB),NOSPLIT,$0
 	MOVD dst_base+0(FP), R0
@@ -29,37 +37,37 @@ TEXT ·encodeAsm(SB),NOSPLIT,$0
 	EOR R5, R5, R5
 
 loop48:
-	CMP $48, R2
-	BLT lessThan48
-
-	// Move the input bits to where they need to be in the outputs. Except
-	// for the first output, the high two bits are not cleared.
-	VLD3.P 48(R1), [V0.B16, V1.B16, V2.B16]
-	VUSHR $2, V0.B16, V3.B16
-	VUSHR $4, V1.B16, V4.B16
-	VUSHR $6, V2.B16, V5.B16
-	VSLI $4, V0.B16, V4.B16
-	VSLI $2, V1.B16, V5.B16
-
-	// Clear the high two bits in the second, third and fourth output.
-	VAND V7.B16, V4.B16, V4.B16
-	VAND V7.B16, V5.B16, V5.B16
-	VAND V7.B16, V2.B16, V6.B16
-
-	// The bits have now been shifted to the right locations;
-	// translate their values 0..63 to the Base64 alphabet.
-	// Use a 64-byte table lookup:
-	VTBL V3.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V3.B16
-	VTBL V4.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V4.B16
-	VTBL V5.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V5.B16
-	VTBL V6.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V6.B16
-
-	// Interleave and store output:
-	VST4.P [V3.B16, V4.B16, V5.B16, V6.B16], 64(R0)
-
-	SUB $48, R2
-	ADD $64, R5
-	B loop48
+		CMP $48, R2
+		BLT lessThan48
+
+		// Move the input bits to where they need to be in the outputs. Except
+		// for the first output, the high two bits are not cleared.
+		VLD3.P 48(R1), [V0.B16, V1.B16, V2.B16]
+		VUSHR $2, V0.B16, V3.B16
+		VUSHR $4, V1.B16, V4.B16
+		VUSHR $6, V2.B16, V5.B16
+		VSLI $4, V0.B16, V4.B16
+		VSLI $2, V1.B16, V5.B16
+
+		// Clear the high two bits in the second, third and fourth output.
+		VAND V7.B16, V4.B16, V4.B16
+		VAND V7.B16, V5.B16, V5.B16
+		VAND V7.B16, V2.B16, V6.B16
+
+		// The bits have now been shifted to the right locations;
+		// translate their values 0..63 to the Base64 alphabet.
+		// Use a 64-byte table lookup:
+		VTBL V3.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V3.B16
+		VTBL V4.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V4.B16
+		VTBL V5.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V5.B16
+		VTBL V6.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V6.B16
+
+		// Interleave and store output:
+		VST4.P [V3.B16, V4.B16, V5.B16, V6.B16], 64(R0)
+
+		SUB $48, R2
+		ADD $64, R5
+		B loop48
 
 lessThan48:
 	// fast return
@@ -73,29 +81,29 @@ lessThan48:
 	VSHL $2, V7.S4, V12.S4 // mulhi constant
 
 loop12:
-	VLD1 (R1), [V0.B16]
-	VTBL V3.B16, [V0.B16], V0.B16 // shuffle bytes
-	VAND V4.B16, V0.B16, V1.B16   // AND mulhi mask
-
-	WORD $0x2e61c182 // UMULL V1.H8, V12.H8, V2.H8
-	WORD $0x6e61c181 // UMULL2 V1.H8, V12.H8, V1.H8
-	VTBL V6.B16, [V1.B16, V2.B16], V1.B16
-
-	VAND V0.B16, V5.B16, V0.B16
-	WORD $0x4e609ce0 // VMUL V0.H8, V7.H8, V0.H8
-	VORR V0.B16, V1.B16, V0.B16
-
-	// The bits have now been shifted to the right locations;
-	// translate their values 0..63 to the Base64 alphabet.
-	// Use a 64-byte table lookup:
-	VTBL V0.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V0.B16
-	VST1.P [V0.B16], 16(R0)
-
-	ADD $12, R1
-	ADD $16, R5
-	SUB $12, R2
-	CMP $16, R2
-	BGE loop12
+		VLD1 (R1), [V0.B16]
+		VTBL V3.B16, [V0.B16], V0.B16 // shuffle bytes
+		VAND V4.B16, V0.B16, V1.B16   // AND mulhi mask
+
+		WORD $0x2e61c182 // UMULL V1.H8, V12.H8, V2.H8
+		WORD $0x6e61c181 // UMULL2 V1.H8, V12.H8, V1.H8
+		VTBL V6.B16, [V1.B16, V2.B16], V1.B16
+
+		VAND V0.B16, V5.B16, V0.B16
+		WORD $0x4e609ce0 // VMUL V0.H8, V7.H8, V0.H8
+		VORR V0.B16, V1.B16, V0.B16
+
+		// The bits have now been shifted to the right locations;
+		// translate their values 0..63 to the Base64 alphabet.
+		// Use a 64-byte table lookup:
+		VTBL V0.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V0.B16
+		VST1.P [V0.B16], 16(R0)
+
+		ADD $12, R1
+		ADD $16, R5
+		SUB $12, R2
+		CMP $16, R2
+		BGE loop12
 
 done:
 	MOVD R5, ret+56(FP)
@@ -113,68 +121,118 @@ TEXT ·decodeAsm(SB),NOSPLIT,$0
 	MOVD $63, R4
 	VDUP R4, V7.B16
 
-loop:
-	CMP $64, R2
-	BLT done
+loop64:
+		CMP $64, R2
+		BLT lessThan64
 
-	VLD4.P 64(R1), [V0.B16, V1.B16, V2.B16, V3.B16]
+		VLD4.P 64(R1), [V0.B16, V1.B16, V2.B16, V3.B16]
 	
-	// Get indices for second LUT:
-	WORD $0x6e272c10 // VUQSUB V7.B16, V0.B16, V16.B16
-	WORD $0x6e272c31 // VUQSUB V7.B16, V1.B16, V17.B16
-	WORD $0x6e272c52 // VUQSUB V7.B16, V2.B16, V18.B16
-	WORD $0x6e272c73 // VUQSUB V7.B16, V3.B16, V19.B16
-
-	// Get values from first LUT:
-	VTBL V0.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V20.B16
-	VTBL V1.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V21.B16
-	VTBL V2.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V22.B16
-	VTBL V3.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V23.B16
-
-	// Get values from second LUT:
-	WORD $0x4e107190 // VTBX V16.B16, [V12.B16, V13.B16, V14.B16, V15.B16], V16.B16
-	WORD $0x4e117191 // VTBX V17.B16, [V12.B16, V13.B16, V14.B16, V15.B16], V17.B16
-	WORD $0x4e127192 // VTBX V18.B16, [V12.B16, V13.B16, V14.B16, V15.B16], V18.B16
-	WORD $0x4e137193 // VTBX V19.B16, [V12.B16, V13.B16, V14.B16, V15.B16], V19.B16
-
-	// Get final values:
-	VORR V16.B16, V20.B16, V0.B16
-	VORR V17.B16, V21.B16, V1.B16
-	VORR V18.B16, V22.B16, V2.B16
-	VORR V19.B16, V23.B16, V3.B16
-
-	// Check for invalid input, any value larger than 63:
-	WORD $0x6e273410 // VCMHI V7.B16, V0.B16, V16.B16
-	WORD $0x6e273431 // VCMHI V7.B16, V1.B16, V17.B16
-	WORD $0x6e273452 // VCMHI V7.B16, V2.B16, V18.B16
-	WORD $0x6e273473 // VCMHI V7.B16, V3.B16, V19.B16
-
-	VORR V17.B16, V16.B16, V16.B16
-	VORR V18.B16, V16.B16, V16.B16
-	VORR V19.B16, V16.B16, V16.B16
-
-	// Check that all bits are zero:
-	WORD $0x6e30aa11 // VUMAXV V16.B16, V17
-	VMOV V17.B[0], R5
-	CBNZ R5, done
-
-	// Compress four bytes into three:
-	VSHL $2, V0.B16, V4.B16
-	VUSHR $4, V1.B16, V16.B16
-	VORR  V16.B16, V4.B16, V4.B16
-
-	VSHL $4, V1.B16, V5.B16
-	VUSHR $2, V2.B16, V16.B16
-	VORR  V16.B16, V5.B16, V5.B16
+		// Get indices for second LUT
+		WORD $0x6e272c10 // VUQSUB V7.B16, V0.B16, V16.B16
+		WORD $0x6e272c31 // VUQSUB V7.B16, V1.B16, V17.B16
+		WORD $0x6e272c52 // VUQSUB V7.B16, V2.B16, V18.B16
+		WORD $0x6e272c73 // VUQSUB V7.B16, V3.B16, V19.B16
+
+		// Get values from first LUT
+		VTBL V0.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V20.B16
+		VTBL V1.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V21.B16
+		VTBL V2.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V22.B16
+		VTBL V3.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V23.B16
+
+		// Get values from second LUT
+		WORD $0x4e107190 // VTBX V16.B16, [V12.B16, V13.B16, V14.B16, V15.B16], V16.B16
+		WORD $0x4e117191 // VTBX V17.B16, [V12.B16, V13.B16, V14.B16, V15.B16], V17.B16
+		WORD $0x4e127192 // VTBX V18.B16, [V12.B16, V13.B16, V14.B16, V15.B16], V18.B16
+		WORD $0x4e137193 // VTBX V19.B16, [V12.B16, V13.B16, V14.B16, V15.B16], V19.B16
+
+		// Get final values:
+		VORR V16.B16, V20.B16, V0.B16
+		VORR V17.B16, V21.B16, V1.B16
+		VORR V18.B16, V22.B16, V2.B16
+		VORR V19.B16, V23.B16, V3.B16
+
+		// Check for invalid input, any value larger than 63
+		WORD $0x6e273410 // VCMHI V7.B16, V0.B16, V16.B16
+		WORD $0x6e273431 // VCMHI V7.B16, V1.B16, V17.B16
+		WORD $0x6e273452 // VCMHI V7.B16, V2.B16, V18.B16
+		WORD $0x6e273473 // VCMHI V7.B16, V3.B16, V19.B16
+
+		VORR V17.B16, V16.B16, V16.B16
+		VORR V18.B16, V16.B16, V16.B16
+		VORR V19.B16, V16.B16, V16.B16
+
+		// Check that all bits are zero:
+		WORD $0x6e30aa11 // VUMAXV V16.B16, V17
+		VMOV V17.B[0], R5
+		CBNZ R5, done
+
+		// Compress four bytes into three
+		VSHL $2, V0.B16, V4.B16
+		VUSHR $4, V1.B16, V16.B16
+		VORR  V16.B16, V4.B16, V4.B16
+
+		VSHL $4, V1.B16, V5.B16
+		VUSHR $2, V2.B16, V16.B16
+		VORR  V16.B16, V5.B16, V5.B16
 	
-	VSHL $6, V2.B16, V16.B16
-	VORR  V16.B16, V3.B16, V6.B16
+		VSHL $6, V2.B16, V16.B16
+		VORR  V16.B16, V3.B16, V6.B16
+
+		// Interleave and store decoded result
+		VST3.P [V4.B16, V5.B16, V6.B16], 48(R0)
+
+		SUB $64, R2
+		B loop64
+
+lessThan64:
+	// fast return
+	CMP $24, R2
+	BLT done
+
+	MOVD $dec_const<>(SB), R4
+	VLD1 (R4), [V1.B16, V2.B16, V3.B16]
 
-	// Interleave and store decoded result:
-	VST3.P [V4.B16, V5.B16, V6.B16], 48(R0)
+loop16:
+		VLD1.P 16(R1), [V0.B16]
 
-	SUB $64, R2
-	B loop
+		// Get indices for second LUT
+		WORD $0x6e272c10 // VUQSUB V7.B16, V0.B16, V16.B16
+
+		// Get values from first LUT
+		VTBL V0.B16, [V8.B16, V9.B16, V10.B16, V11.B16], V20.B16
+	
+		// Get values from second LUT
+		WORD $0x4e107190 // VTBX V16.B16, [V12.B16, V13.B16, V14.B16, V15.B16], V16.B16
+
+		// Get final values:
+		VORR V16.B16, V20.B16, V0.B16
+
+		// Check for invalid input, any value larger than 63
+		WORD $0x6e273410 // VCMHI V7.B16, V0.B16, V16.B16
+		// Check that all bits are zero:
+		WORD $0x6e30aa11 // VUMAXV V16.B16, V17
+		VMOV V17.B[0], R5
+		CBNZ R5, done
+	
+		// Compress four bytes into three
+		// swap and merge adjacent 6-bit fields
+		WORD $0x2e20c024 // UMULL V0.B16, V1.B16, V4.H8
+		WORD $0x6e20c020 // UMULL2 V0.B16, V1.B16, V0.H8
+		VADDP V0.H8, V4.H8, V0.H8
+
+		// swap and merge 12-bit words into a 24-bit word
+		WORD $0x2e60c044 // UMULL V0.H8, V2.H8, V4.S4
+		WORD $0x6e60c040 // UMULL2 V0.H8, V2.H8, V0.S4
+		VADDP V0.S4, V4.S4, V0.S4
+
+		// reshuffle bytes
+		VTBL V3.B16, [V0.B16], V0.B16
+		VST1 [V0.B16], (R0)
+
+		ADD $12, R0
+		SUB $16, R2
+		CMP $24, R2
+		BGE loop16
 
 done:
 	MOVD R2, ret+56(FP)
diff --git a/base64_arm64_test.go b/base64_arm64_test.go
index 2ab87de..ffc79e8 100644
--- a/base64_arm64_test.go
+++ b/base64_arm64_test.go
@@ -35,21 +35,24 @@ func TestStdEncodeSIMD(t *testing.T) {
 
 func TestStdDecodeSIMD(t *testing.T) {
 	pairs := []testpair{
+		{"abcdefghijkl", "YWJjZGVmZ2hpamtsYWJjZGVmZ2hpamts"},
+		{"\x2b\xf7\xcc\x27\x01\xfe\x43\x97\xb4\x9e\xbe\xed", "K/fMJwH+Q5e0nr7tK/fMJwH+Q5e0nr7t"},
+		{"abcdefghijklabcdefghijklabcdefghijkl", "YWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamts"},
 		{"abcdefghijklabcdefghijklabcdefghijklabcdefghijkl", "YWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamts"},
 		{"abcdefghijklabcdefghijklabcdefghijklabcdefghijkl", "YWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamt="},
 		{"abcdefghijklabcdefghijklabcdefghijklabcdefghijklabcdefghijklabcdefghijklabcdefghijklabcdefghijkl", "YWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamts"},
 	}
-	for _, p := range pairs {
+	for i, p := range pairs {
 		expected := []byte(p.decoded)
 		src := []byte(p.encoded)
 		dst := make([]byte, len(expected))
 
 		ret := decodeAsm(dst, src, &dencodeStdLut)
 		if ret == len(src) {
-			t.Fatal("should return decode")
+			t.Fatalf("case %v should return decode %v, got %v", i, len(src), ret)
 		}
 		if !bytes.Equal(dst, expected) {
-			t.Fatalf("got %x, expected %x", dst, expected)
+			t.Fatalf("case %v got %x, expected %x", i, dst, expected)
 		}
 	}
 }
@@ -58,7 +61,7 @@ func TestUrlEncodeSIMD(t *testing.T) {
 	pairs := []testpair{
 		{"!?$*&()'-=@~0000", "IT8kKiYoKSctPUB-"},
 		{"\x2b\xf7\xcc\x27\x01\xfe\x43\x97\xb4\x9e\xbe\xed\x5a\xcc\x70\x90", "K_fMJwH-Q5e0nr7t"},
-		{"!?$*&()'-=@~!?$*&()'-=@~0000", "IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-"},		
+		{"!?$*&()'-=@~!?$*&()'-=@~0000", "IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-"},
 		{"!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~", "IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-"},
 		{"!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~", "IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-"},
 	}
@@ -80,21 +83,24 @@ func TestUrlEncodeSIMD(t *testing.T) {
 
 func TestUrlDecodeSIMD(t *testing.T) {
 	pairs := []testpair{
+		{"!?$*&()'-=@~", "IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-"},
+		{"\x2b\xf7\xcc\x27\x01\xfe\x43\x97\xb4\x9e\xbe\xed", "K_fMJwH-Q5e0nr7tK_fMJwH-Q5e0nr7t"},
+		{"!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~", "IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-"},
 		{"!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~", "IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-"},
 		{"abcdefghijklabcdefghijklabcdefghijklabcdefghijkl", "YWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamtsYWJjZGVmZ2hpamt="},
 		{"!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~!?$*&()'-=@~", "IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-IT8kKiYoKSctPUB-"},
 	}
-	for _, p := range pairs {
+	for i, p := range pairs {
 		expected := []byte(p.decoded)
 		src := []byte(p.encoded)
 		dst := make([]byte, len(expected))
 
 		ret := decodeAsm(dst, src, &dencodeUrlLut)
 		if ret == len(src) {
-			t.Fatal("should return decode")
+			t.Fatalf("case %v should return decode %v, got %v", i, len(src), ret)
 		}
 		if !bytes.Equal(dst, expected) {
-			t.Fatalf("got %x, expected %x", dst, expected)
+			t.Fatalf("case %v got %x, expected %x", i, dst, expected)
 		}
 	}
 }