diff --git a/include/libbase64.h b/include/libbase64.h
index 16c8ef0e..c7ec0b2c 100644
--- a/include/libbase64.h
+++ b/include/libbase64.h
@@ -1,6 +1,8 @@
 #ifndef _LIBBASE64_H
 #define _LIBBASE64_H
 
+#include <stdint.h>
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -80,6 +82,14 @@ int base64_decode
 	, int			 flags
 	) ;
 
+int base64_decode16
+	( const uint16_t	*src
+	, size_t			 srclen
+	, char				*out
+	, size_t			*outlen
+	, int				 flags
+	) ;
+
 /* Call this before calling base64_stream_decode() to init the state. See above
  * for `flags`; set to 0 for default operation: */
 void base64_stream_decode_init
@@ -103,6 +113,14 @@ int base64_stream_decode
 	, size_t		*outlen
 	) ;
 
+int base64_stream_decode16
+	( struct base64_state	*state
+	, const uint16_t		*src
+	, size_t				 srclen
+	, char					*out
+	, size_t				*outlen
+	) ;
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/lib/arch/avx/codec.c b/lib/arch/avx/codec.c
index d1a6164f..04e8b64e 100644
--- a/lib/arch/avx/codec.c
+++ b/lib/arch/avx/codec.c
@@ -15,6 +15,8 @@
 #include "../ssse3/enc_translate.c"
 #include "../sse41/enc_reshuffle.c"
 
+#include "../generic/convert.c"
+
 #endif	// __AVX__
 
 BASE64_ENC_FUNCTION(avx)
@@ -28,6 +30,9 @@ BASE64_ENC_FUNCTION(avx)
 #endif
 }
 
+#define STRING_TYPE uint8_t
+#define CHAR_CONVERT(x) (x)
+#define LOAD_STRING(c) LOAD_STRING8(c)
 BASE64_DEC_FUNCTION(avx)
 {
 #ifdef __AVX__
@@ -38,3 +43,33 @@ BASE64_DEC_FUNCTION(avx)
 	BASE64_DEC_STUB
 #endif
 }
+#undef LOAD_STRING
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+#define STRING_TYPE uint16_t
+#define CHAR_CONVERT(x) convert(x)
+#define LOAD_STRING(c) LOAD_STRING16(c)
+BASE64_DEC16_FUNCTION(avx)
+{
+#ifdef __AVX__
+#include "../generic/dec_head.c"
+#include "../sse42/dec_loop.c"
+#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}
+#undef LOAD_STRING
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+BASE64_CVT_FUNCTION(avx)
+{
+#ifdef __AVX__
+	#include "../sse2/convert_loop.c"
+	#include "../generic/convert_loop.c"
+#else
+	BASE64_CVT_STUB
+#endif
+}
diff --git a/lib/arch/avx2/codec.c b/lib/arch/avx2/codec.c
index 114a2bb9..20448549 100644
--- a/lib/arch/avx2/codec.c
+++ b/lib/arch/avx2/codec.c
@@ -130,6 +130,8 @@ dec_reshuffle (__m256i in)
 		0, 1, 2, 4, 5, 6, -1, -1));
 }
 
+#include "../generic/convert.c"
+
 #endif	// __AVX2__
 
 BASE64_ENC_FUNCTION(avx2)
@@ -143,6 +145,9 @@ BASE64_ENC_FUNCTION(avx2)
 #endif
 }
 
+#define STRING_TYPE uint8_t
+#define CHAR_CONVERT(x) (x)
+#define LOAD_STRING(c) _mm256_loadu_si256((__m256i *)c)
 BASE64_DEC_FUNCTION(avx2)
 {
 #ifdef __AVX2__
@@ -153,3 +158,33 @@ BASE64_DEC_FUNCTION(avx2)
 	BASE64_DEC_STUB
 #endif
 }
+#undef LOAD_STRING
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+#define STRING_TYPE uint16_t
+#define CHAR_CONVERT(x) convert(x)
+#define LOAD_STRING(c) _mm256_permute4x64_epi64(_mm256_packus_epi16( _mm256_loadu_si256((__m256i *)(c+0)),_mm256_loadu_si256((__m256i *)(c+16))), 0xD8)
+BASE64_DEC16_FUNCTION(avx2)
+{
+#ifdef __AVX2__
+#include "../generic/dec_head.c"
+#include "dec_loop.c"
+#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}
+#undef LOAD_STRING
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+BASE64_CVT_FUNCTION(avx2)
+{
+#ifdef __AVX2__
+	#include "../sse2/convert_loop.c"
+	#include "../generic/convert_loop.c"
+#else
+	BASE64_CVT_STUB
+#endif
+}
diff --git a/lib/arch/avx2/dec_loop.c b/lib/arch/avx2/dec_loop.c
index 8344004c..5765ef03 100644
--- a/lib/arch/avx2/dec_loop.c
+++ b/lib/arch/avx2/dec_loop.c
@@ -6,7 +6,7 @@
 while (srclen >= 45)
 {
 	// Load string:
-	__m256i str = _mm256_loadu_si256((__m256i *)c);
+	__m256i str = LOAD_STRING(c);
 
 	// The input consists of six character sets in the Base64 alphabet,
 	// which we need to map back to the 6-bit values they represent.
diff --git a/lib/arch/generic/codec.c b/lib/arch/generic/codec.c
index ae54a8b8..4b3464a1 100644
--- a/lib/arch/generic/codec.c
+++ b/lib/arch/generic/codec.c
@@ -5,6 +5,8 @@
 #include "../../../include/libbase64.h"
 #include "../../codecs.h"
 
+#include "../generic/convert.c"
+
 BASE64_ENC_FUNCTION(plain)
 {
 	#include "enc_head.c"
@@ -16,6 +18,8 @@ BASE64_ENC_FUNCTION(plain)
 	#include "enc_tail.c"
 }
 
+#define STRING_TYPE uint8_t
+#define CHAR_CONVERT(x) (x)
 BASE64_DEC_FUNCTION(plain)
 {
 	#include "dec_head.c"
@@ -26,3 +30,20 @@ BASE64_DEC_FUNCTION(plain)
 #endif
 	#include "dec_tail.c"
 }
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+#define STRING_TYPE uint16_t
+#define CHAR_CONVERT(x) convert(x)
+BASE64_DEC16_FUNCTION(plain)
+{
+#include "dec_head.c"
+#include "dec_tail.c"
+}
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+BASE64_CVT_FUNCTION(plain)
+{
+	#include "convert_loop.c"
+}
diff --git a/lib/arch/generic/convert.c b/lib/arch/generic/convert.c
new file mode 100644
index 00000000..7400b7d9
--- /dev/null
+++ b/lib/arch/generic/convert.c
@@ -0,0 +1,13 @@
+static inline uint8_t
+convert (const uint16_t in)
+{
+	unsigned value = in;
+#if 0
+	unsigned mask = (unsigned)((0 - (int)(value >> 8)) >> (sizeof(int) * 8U - 1U));
+
+	value |= mask;
+	return (uint8_t)value;
+#else
+	return (value > 255U) ? 255U : value;
+#endif
+}
diff --git a/lib/arch/generic/convert_loop.c b/lib/arch/generic/convert_loop.c
new file mode 100644
index 00000000..b606ebd1
--- /dev/null
+++ b/lib/arch/generic/convert_loop.c
@@ -0,0 +1,4 @@
+while (len > 0U) {
+	*dst++ = (char)convert(*src++);
+	len--;
+}
diff --git a/lib/arch/generic/dec_head.c b/lib/arch/generic/dec_head.c
index 4aedcd3f..425fcdb2 100644
--- a/lib/arch/generic/dec_head.c
+++ b/lib/arch/generic/dec_head.c
@@ -1,5 +1,5 @@
 int ret = 0;
-const uint8_t *c = (const uint8_t *)src;
+const STRING_TYPE *c = (const STRING_TYPE *)src;
 uint8_t *o = (uint8_t *)out;
 uint8_t q;
 
@@ -17,7 +17,7 @@ if (st.eof) {
 	// If there was a trailing '=' to check, check it:
 	if (srclen && (st.eof == BASE64_AEOF)) {
 		st.eof = BASE64_EOF;
-		ret = (base64_table_dec[*c++] == 254) ? 1 : 0;
+		ret = (base64_table_dec[CHAR_CONVERT(*c++)] == 254) ? 1 : 0;
 	}
 	return ret;
 }
diff --git a/lib/arch/generic/dec_tail.c b/lib/arch/generic/dec_tail.c
index be40fedf..03a9c130 100644
--- a/lib/arch/generic/dec_tail.c
+++ b/lib/arch/generic/dec_tail.c
@@ -2,7 +2,7 @@
 			ret = 1;
 			break;
 		}
-		if ((q = base64_table_dec[*c++]) >= 254) {
+		if ((q = base64_table_dec[CHAR_CONVERT(*c++)]) >= 254) {
 			st.eof = BASE64_EOF;
 			// Treat character '=' as invalid for byte 0:
 			break;
@@ -14,7 +14,7 @@
 			ret = 1;
 			break;
 		}
-		if ((q = base64_table_dec[*c++]) >= 254) {
+		if ((q = base64_table_dec[CHAR_CONVERT(*c++)]) >= 254) {
 			st.eof = BASE64_EOF;
 			// Treat character '=' as invalid for byte 1:
 			break;
@@ -28,14 +28,14 @@
 			ret = 1;
 			break;
 		}
-		if ((q = base64_table_dec[*c++]) >= 254) {
+		if ((q = base64_table_dec[CHAR_CONVERT(*c++)]) >= 254) {
 			// When q == 254, the input char is '='.
 			// Check if next byte is also '=':
 			if (q == 254) {
 				if (srclen-- != 0) {
 					// EOF:
 					st.eof = BASE64_EOF;
-					ret = (base64_table_dec[*c++] == 254) ? 1 : 0;
+					ret = (base64_table_dec[CHAR_CONVERT(*c++)] == 254) ? 1 : 0;
 					break;
 				}
 				else {
@@ -58,7 +58,7 @@
 			ret = 1;
 			break;
 		}
-		if ((q = base64_table_dec[*c++]) >= 254) {
+		if ((q = base64_table_dec[CHAR_CONVERT(*c++)]) >= 254) {
 			st.eof = BASE64_EOF;
 			// When q == 254, the input char is '='. Return 1 and EOF.
 			// When q == 255, the input char is invalid. Return 0 and EOF.
diff --git a/lib/arch/neon32/codec.c b/lib/arch/neon32/codec.c
index e7288fcc..25f0a579 100644
--- a/lib/arch/neon32/codec.c
+++ b/lib/arch/neon32/codec.c
@@ -111,6 +111,8 @@ enc_translate (uint8x16x4_t in)
 	return out;
 }
 
+#include "../generic/convert.c"
+
 #endif
 
 // Stride size is so large on these NEON 32-bit functions
@@ -140,3 +142,13 @@ BASE64_DEC_FUNCTION(neon32)
 	BASE64_DEC_STUB
 #endif
 }
+
+
+BASE64_CVT_FUNCTION(neon32)
+{
+#if (defined(__arm__) && defined(__ARM_NEON__))
+	#include "../generic/convert_loop.c"
+#else
+	BASE64_CVT_STUB
+#endif
+}
diff --git a/lib/arch/neon64/codec.c b/lib/arch/neon64/codec.c
index c101df7c..cece1dbf 100644
--- a/lib/arch/neon64/codec.c
+++ b/lib/arch/neon64/codec.c
@@ -41,6 +41,9 @@ static const char *base64_table_enc_transposed =
 	"Oeu+"
 	"Pfv/"
 };
+
+#include "../generic/convert.c"
+
 #endif
 
 // Stride size is so large on these NEON 64-bit functions
@@ -73,3 +76,12 @@ BASE64_DEC_FUNCTION(neon64)
 	BASE64_DEC_STUB
 #endif
 }
+
+BASE64_CVT_FUNCTION(neon64)
+{
+#if (defined(__aarch64__) && defined(__ARM_NEON__))
+	#include "../generic/convert_loop.c"
+#else
+	BASE64_CVT_STUB
+#endif
+}
diff --git a/lib/arch/sse2/compare_macros.h b/lib/arch/sse2/compare_macros.h
index 76d92887..d06e1218 100644
--- a/lib/arch/sse2/compare_macros.h
+++ b/lib/arch/sse2/compare_macros.h
@@ -2,3 +2,5 @@
 #define CMPEQ(s,n)	_mm_cmpeq_epi8((s), _mm_set1_epi8(n))
 #define REPLACE(s,n)	_mm_and_si128((s), _mm_set1_epi8(n))
 #define RANGE(s,a,b)	_mm_andnot_si128(CMPGT((s), (b)), CMPGT((s), (a) - 1))
+#define LOAD_STRING8(c) _mm_loadu_si128((__m128i *)c)
+#define LOAD_STRING16(c) _mm_packus_epi16(_mm_loadu_si128((__m128i *)(c + 0)), _mm_loadu_si128((__m128i *)(c + 8)))
diff --git a/lib/arch/sse2/convert_loop.c b/lib/arch/sse2/convert_loop.c
new file mode 100644
index 00000000..f9c2537d
--- /dev/null
+++ b/lib/arch/sse2/convert_loop.c
@@ -0,0 +1,53 @@
+while (len >= 64U) {
+	__m128i src0 = _mm_loadu_si128((const __m128i*)(src +  0));
+	__m128i src1 = _mm_loadu_si128((const __m128i*)(src +  8));
+	__m128i src2 = _mm_loadu_si128((const __m128i*)(src + 16));
+	__m128i src3 = _mm_loadu_si128((const __m128i*)(src + 24));
+	__m128i src4 = _mm_loadu_si128((const __m128i*)(src + 32));
+	__m128i src5 = _mm_loadu_si128((const __m128i*)(src + 40));
+	__m128i src6 = _mm_loadu_si128((const __m128i*)(src + 48));
+	__m128i src7 = _mm_loadu_si128((const __m128i*)(src + 56));
+
+
+	__m128i dst0 = _mm_packus_epi16(src0, src1);
+	__m128i dst1 = _mm_packus_epi16(src2, src3);
+	__m128i dst2 = _mm_packus_epi16(src4, src5);
+	__m128i dst3 = _mm_packus_epi16(src6, src7);
+
+	_mm_storeu_si128((__m128i*)(dst +  0), dst0);
+	_mm_storeu_si128((__m128i*)(dst + 16), dst1);
+	_mm_storeu_si128((__m128i*)(dst + 32), dst2);
+	_mm_storeu_si128((__m128i*)(dst + 48), dst3);
+
+	len-= 64U;
+	src += 64U;
+	dst += 64U;
+}
+if (len & 32U) {
+	__m128i src0 = _mm_loadu_si128((const __m128i*)(src +  0));
+	__m128i src1 = _mm_loadu_si128((const __m128i*)(src +  8));
+	__m128i src2 = _mm_loadu_si128((const __m128i*)(src + 16));
+	__m128i src3 = _mm_loadu_si128((const __m128i*)(src + 24));
+
+	__m128i dst0 = _mm_packus_epi16(src0, src1);
+	__m128i dst1 = _mm_packus_epi16(src2, src3);
+
+	_mm_storeu_si128((__m128i*)(dst +  0), dst0);
+	_mm_storeu_si128((__m128i*)(dst + 16), dst1);
+
+	len-= 32U;
+	src += 32U;
+	dst += 32U;
+}
+if (len & 16U) {
+	__m128i src0 = _mm_loadu_si128((const __m128i*)(src +  0));
+	__m128i src1 = _mm_loadu_si128((const __m128i*)(src +  8));
+
+	__m128i dst0 = _mm_packus_epi16(src0, src1);
+
+	_mm_storeu_si128((__m128i*)(dst +  0), dst0);
+
+	len-= 16U;
+	src += 16U;
+	dst += 16U;
+}
diff --git a/lib/arch/sse41/codec.c b/lib/arch/sse41/codec.c
index d9e33b5e..20526072 100644
--- a/lib/arch/sse41/codec.c
+++ b/lib/arch/sse41/codec.c
@@ -15,6 +15,8 @@
 #include "../ssse3/enc_translate.c"
 #include "enc_reshuffle.c"
 
+#include "../generic/convert.c"
+
 #endif	// __SSE4_1__
 
 BASE64_ENC_FUNCTION(sse41)
@@ -28,6 +30,9 @@ BASE64_ENC_FUNCTION(sse41)
 #endif
 }
 
+#define STRING_TYPE uint8_t
+#define CHAR_CONVERT(x) (x)
+#define LOAD_STRING(c) LOAD_STRING8(c)
 BASE64_DEC_FUNCTION(sse41)
 {
 #ifdef __SSE4_1__
@@ -38,3 +43,33 @@ BASE64_DEC_FUNCTION(sse41)
 	BASE64_DEC_STUB
 #endif
 }
+#undef LOAD_STRING
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+#define STRING_TYPE uint16_t
+#define CHAR_CONVERT(x) convert(x)
+#define LOAD_STRING(c) LOAD_STRING16(c)
+BASE64_DEC16_FUNCTION(sse41)
+{
+#ifdef __SSE4_1__
+#include "../generic/dec_head.c"
+#include "../ssse3/dec_loop.c"
+#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}
+#undef LOAD_STRING
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+BASE64_CVT_FUNCTION(sse41)
+{
+#ifdef __SSE4_1__
+	#include "../sse2/convert_loop.c"
+	#include "../generic/convert_loop.c"
+#else
+	BASE64_CVT_STUB
+#endif
+}
diff --git a/lib/arch/sse42/codec.c b/lib/arch/sse42/codec.c
index e9c1e6ca..c1671739 100644
--- a/lib/arch/sse42/codec.c
+++ b/lib/arch/sse42/codec.c
@@ -15,6 +15,8 @@
 #include "../ssse3/enc_translate.c"
 #include "../sse41/enc_reshuffle.c"
 
+#include "../generic/convert.c"
+
 #endif	// __SSE4_2__
 
 BASE64_ENC_FUNCTION(sse42)
@@ -28,6 +30,9 @@ BASE64_ENC_FUNCTION(sse42)
 #endif
 }
 
+#define STRING_TYPE uint8_t
+#define CHAR_CONVERT(x) (x)
+#define LOAD_STRING(c) LOAD_STRING8(c)
 BASE64_DEC_FUNCTION(sse42)
 {
 #ifdef __SSE4_2__
@@ -38,3 +43,33 @@ BASE64_DEC_FUNCTION(sse42)
 	BASE64_DEC_STUB
 #endif
 }
+#undef LOAD_STRING
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+#define STRING_TYPE uint16_t
+#define CHAR_CONVERT(x) convert(x)
+#define LOAD_STRING(c) LOAD_STRING16(c)
+BASE64_DEC16_FUNCTION(sse42)
+{
+#ifdef __SSE4_2__
+#include "../generic/dec_head.c"
+#include "dec_loop.c"
+#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}
+#undef LOAD_STRING
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+BASE64_CVT_FUNCTION(sse42)
+{
+#ifdef __SSE4_2__
+	#include "../sse2/convert_loop.c"
+	#include "../generic/convert_loop.c"
+#else
+	BASE64_CVT_STUB
+#endif
+}
diff --git a/lib/arch/sse42/dec_loop.c b/lib/arch/sse42/dec_loop.c
index f7e55dcf..a7106ad3 100644
--- a/lib/arch/sse42/dec_loop.c
+++ b/lib/arch/sse42/dec_loop.c
@@ -6,7 +6,7 @@
 while (srclen >= 24)
 {
 	// Load string:
-	__m128i str = _mm_loadu_si128((__m128i *)c);
+	__m128i str = LOAD_STRING(c);
 
 	// The input consists of six character sets in the Base64 alphabet,
 	// which we need to map back to the 6-bit values they represent.
diff --git a/lib/arch/ssse3/codec.c b/lib/arch/ssse3/codec.c
index b00cb0af..dd2adbca 100644
--- a/lib/arch/ssse3/codec.c
+++ b/lib/arch/ssse3/codec.c
@@ -49,6 +49,8 @@ enc_reshuffle (__m128i in)
 	return _mm_bswap_epi32(indices);
 }
 
+#include "../generic/convert.c"
+
 #endif	// __SSSE3__
 
 BASE64_ENC_FUNCTION(ssse3)
@@ -62,6 +64,9 @@ BASE64_ENC_FUNCTION(ssse3)
 #endif
 }
 
+#define STRING_TYPE uint8_t
+#define CHAR_CONVERT(x) (x)
+#define LOAD_STRING(c) LOAD_STRING8(c)
 BASE64_DEC_FUNCTION(ssse3)
 {
 #ifdef __SSSE3__
@@ -72,3 +77,33 @@ BASE64_DEC_FUNCTION(ssse3)
 	BASE64_DEC_STUB
 #endif
 }
+#undef LOAD_STRING
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+#define STRING_TYPE uint16_t
+#define CHAR_CONVERT(x) convert(x)
+#define LOAD_STRING(c) LOAD_STRING16(c)
+BASE64_DEC16_FUNCTION(ssse3)
+{
+#ifdef __SSSE3__
+#include "../generic/dec_head.c"
+#include "dec_loop.c"
+#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}
+#undef LOAD_STRING
+#undef CHAR_CONVERT
+#undef STRING_TYPE
+
+BASE64_CVT_FUNCTION(ssse3)
+{
+#ifdef __SSSE3__
+	#include "../sse2/convert_loop.c"
+	#include "../generic/convert_loop.c"
+#else
+	BASE64_CVT_STUB
+#endif
+}
diff --git a/lib/arch/ssse3/dec_loop.c b/lib/arch/ssse3/dec_loop.c
index e660d4c6..d7a83409 100644
--- a/lib/arch/ssse3/dec_loop.c
+++ b/lib/arch/ssse3/dec_loop.c
@@ -6,7 +6,7 @@
 while (srclen >= 24)
 {
 	// Load string:
-	__m128i str = _mm_loadu_si128((__m128i *)c);
+	__m128i str = LOAD_STRING(c);
 
 	// The input consists of six character sets in the Base64 alphabet,
 	// which we need to map back to the 6-bit values they represent.
diff --git a/lib/codec_choose.c b/lib/codec_choose.c
index 0180aa5f..5089380f 100644
--- a/lib/codec_choose.c
+++ b/lib/codec_choose.c
@@ -66,15 +66,22 @@
 #define BASE64_CODEC_FUNCS(arch)	\
 	BASE64_ENC_FUNCTION(arch);	\
 	BASE64_DEC_FUNCTION(arch);	\
+	BASE64_CVT_FUNCTION(arch);
 
 BASE64_CODEC_FUNCS(avx2)
+BASE64_DEC16_FUNCTION(avx2);
 BASE64_CODEC_FUNCS(neon32)
 BASE64_CODEC_FUNCS(neon64)
 BASE64_CODEC_FUNCS(plain)
+BASE64_DEC16_FUNCTION(plain);
 BASE64_CODEC_FUNCS(ssse3)
+BASE64_DEC16_FUNCTION(ssse3);
 BASE64_CODEC_FUNCS(sse41)
+BASE64_DEC16_FUNCTION(sse41);
 BASE64_CODEC_FUNCS(sse42)
+BASE64_DEC16_FUNCTION(sse42);
 BASE64_CODEC_FUNCS(avx)
+BASE64_DEC16_FUNCTION(avx);
 
 static bool
 codec_choose_forced (struct codec *codec, int flags)
@@ -86,44 +93,59 @@ codec_choose_forced (struct codec *codec, int flags)
 	if (!(flags & 0xFF)) {
 		return false;
 	}
+	codec->dec16 = NULL;
 	if (flags & BASE64_FORCE_AVX2) {
 		codec->enc = base64_stream_encode_avx2;
 		codec->dec = base64_stream_decode_avx2;
+		codec->dec16 = base64_stream_decode16_avx2;
+		codec->cvt = base64_cvt_avx2;
 		return true;
 	}
 	if (flags & BASE64_FORCE_NEON32) {
 		codec->enc = base64_stream_encode_neon32;
 		codec->dec = base64_stream_decode_neon32;
+		codec->cvt = base64_cvt_neon32;
 		return true;
 	}
 	if (flags & BASE64_FORCE_NEON64) {
 		codec->enc = base64_stream_encode_neon64;
 		codec->dec = base64_stream_decode_neon64;
+		codec->cvt = base64_cvt_neon64;
 		return true;
 	}
 	if (flags & BASE64_FORCE_PLAIN) {
 		codec->enc = base64_stream_encode_plain;
 		codec->dec = base64_stream_decode_plain;
+		codec->dec16 = base64_stream_decode16_plain;
+		codec->cvt = base64_cvt_plain;
 		return true;
 	}
 	if (flags & BASE64_FORCE_SSSE3) {
 		codec->enc = base64_stream_encode_ssse3;
 		codec->dec = base64_stream_decode_ssse3;
+		codec->dec16 = base64_stream_decode16_ssse3;
+		codec->cvt = base64_cvt_ssse3;
 		return true;
 	}
 	if (flags & BASE64_FORCE_SSE41) {
 		codec->enc = base64_stream_encode_sse41;
 		codec->dec = base64_stream_decode_sse41;
+		codec->dec16 = base64_stream_decode16_sse41;
+		codec->cvt = base64_cvt_sse41;
 		return true;
 	}
 	if (flags & BASE64_FORCE_SSE42) {
 		codec->enc = base64_stream_encode_sse42;
 		codec->dec = base64_stream_decode_sse42;
+		codec->dec16 = base64_stream_decode16_sse42;
+		codec->cvt = base64_cvt_sse42;
 		return true;
 	}
 	if (flags & BASE64_FORCE_AVX) {
 		codec->enc = base64_stream_encode_avx;
 		codec->dec = base64_stream_decode_avx;
+		codec->dec16 = base64_stream_decode16_avx;
+		codec->cvt = base64_cvt_avx;
 		return true;
 	}
 	return false;
@@ -141,9 +163,11 @@ codec_choose_arm (struct codec *codec)
 	#if defined(__aarch64__) && HAVE_NEON64
 	codec->enc = base64_stream_encode_neon64;
 	codec->dec = base64_stream_decode_neon64;
+	codec->cvt = base64_cvt_neon64;
 	#else
 	codec->enc = base64_stream_encode_neon32;
 	codec->dec = base64_stream_decode_neon32;
+	codec->cvt = base64_cvt_neon32;
 	#endif
 
 	return true;
@@ -192,6 +216,7 @@ codec_choose_x86 (struct codec *codec)
 				if (ebx & bit_AVX2) {
 					codec->enc = base64_stream_encode_avx2;
 					codec->dec = base64_stream_decode_avx2;
+					codec->cvt = base64_cvt_avx2;
 					return true;
 				}
 				#endif
@@ -199,6 +224,7 @@ codec_choose_x86 (struct codec *codec)
 				if (ecx & bit_AVX) {
 					codec->enc = base64_stream_encode_avx;
 					codec->dec = base64_stream_decode_avx;
+					codec->cvt = base64_cvt_avx;
 					return true;
 				}
 				#endif
@@ -214,6 +240,7 @@ codec_choose_x86 (struct codec *codec)
 		if (ecx & bit_SSE42) {
 			codec->enc = base64_stream_encode_sse42;
 			codec->dec = base64_stream_decode_sse42;
+			codec->cvt = base64_cvt_sse42;
 			return true;
 		}
 	}
@@ -226,6 +253,7 @@ codec_choose_x86 (struct codec *codec)
 		if (ecx & bit_SSE41) {
 			codec->enc = base64_stream_encode_sse41;
 			codec->dec = base64_stream_decode_sse41;
+			codec->cvt = base64_cvt_sse41;
 			return true;
 		}
 	}
@@ -238,6 +266,7 @@ codec_choose_x86 (struct codec *codec)
 		if (ecx & bit_SSSE3) {
 			codec->enc = base64_stream_encode_ssse3;
 			codec->dec = base64_stream_decode_ssse3;
+			codec->cvt = base64_cvt_ssse3;
 			return true;
 		}
 	}
@@ -267,4 +296,5 @@ codec_choose (struct codec *codec, int flags)
 	}
 	codec->enc = base64_stream_encode_plain;
 	codec->dec = base64_stream_decode_plain;
+	codec->cvt = base64_cvt_plain;
 }
diff --git a/lib/codecs.h b/lib/codecs.h
index 11ed7829..9bcc998e 100644
--- a/lib/codecs.h
+++ b/lib/codecs.h
@@ -15,6 +15,20 @@
 	, char			*out		\
 	, size_t		*outlen		\
 	)
+#define BASE64_DEC16_PARAMS		\
+	( struct base64_state	*state	\
+	, const uint16_t	*src		\
+	, size_t		 srclen		\
+	, char			*out		\
+	, size_t		*outlen		\
+	)
+
+// Function parameters for conversion functions:
+#define BASE64_CVT_PARAMS		\
+	( const uint16_t	*src	\
+	, char				*dst	\
+	, size_t			 len	\
+	)
 
 // Function signature for encoding functions:
 #define BASE64_ENC_FUNCTION(arch)		\
@@ -28,6 +42,17 @@
 	base64_stream_decode_ ## arch		\
 	BASE64_DEC_PARAMS
 
+#define BASE64_DEC16_FUNCTION(arch)		\
+	int					\
+	base64_stream_decode16_ ## arch		\
+	BASE64_DEC_PARAMS
+
+// Function signature for decoding functions:
+#define BASE64_CVT_FUNCTION(arch)	\
+	void				\
+	base64_cvt_ ## arch	\
+	BASE64_CVT_PARAMS
+
 // Cast away unused variable, silence compiler:
 #define UNUSED(x)		((void)(x))
 
@@ -50,10 +75,18 @@
 						\
 	return -1;
 
+// Stub function when conversion arch unsupported:
+#define BASE64_CVT_STUB				\
+	UNUSED(src);				\
+	UNUSED(dst);				\
+	UNUSED(len);
+
 struct codec
 {
-	void (* enc) BASE64_ENC_PARAMS;
-	int  (* dec) BASE64_DEC_PARAMS;
+	void (* enc)   BASE64_ENC_PARAMS;
+	int  (* dec)   BASE64_DEC_PARAMS;
+	int  (* dec16) BASE64_DEC16_PARAMS;
+	void (* cvt)   BASE64_CVT_PARAMS;
 };
 
 // Define machine endianness. This is for GCC:
diff --git a/lib/lib.c b/lib/lib.c
index c02e2db9..1abbc87b 100644
--- a/lib/lib.c
+++ b/lib/lib.c
@@ -120,6 +120,44 @@ base64_stream_decode
 	return codec.dec(state, src, srclen, out, outlen);
 }
 
+int
+base64_stream_decode16
+	( struct base64_state	*state
+	, const uint16_t		*src
+	, size_t				 srclen
+	, char					*out
+	, size_t				*outlen
+	)
+{
+	int ret = 1;
+	size_t asciilen = (srclen > (4U * 1024U)) ? (4U * 1024U) : srclen;
+	char ascii[asciilen] __attribute__((aligned(64)));
+	size_t outl = 0U;
+
+	if (0 && codec.dec16) {
+		return codec.dec16(state, src, srclen, out, outlen);
+	}
+
+	while (ret && (asciilen < srclen))
+	{
+		size_t outlenround = 0U;
+		codec.cvt(src, ascii, asciilen);
+		ret = codec.dec(state, ascii, asciilen, out, &outlenround);
+		src += asciilen;
+		srclen -= asciilen;
+		out += outlenround;
+		outl += outlenround;
+	}
+	if (ret) {
+		size_t outlenround = 0U;
+		codec.cvt(src, ascii, srclen);
+		ret = codec.dec(state, ascii, srclen, out, &outlenround);
+		outl += outlenround;
+	}
+	*outlen = outl;
+	return ret;
+}
+
 #ifdef _OPENMP
 
 	// Due to the overhead of initializing OpenMP and creating a team of
@@ -186,3 +224,21 @@ base64_decode
 	// Feed the whole string to the stream reader:
 	return base64_stream_decode(&state, src, srclen, out, outlen);
 }
+
+int base64_decode16
+ ( const uint16_t	*src
+ , size_t			 srclen
+ , char				*out
+ , size_t			*outlen
+ , int				 flags
+ )
+{
+	struct base64_state state;
+
+	// Init the stream reader:
+	base64_stream_decode_init(&state, flags);
+
+	// Feed the whole string to the stream reader:
+	return base64_stream_decode16(&state, src, srclen, out, outlen);
+}
+
diff --git a/test/benchmark.c b/test/benchmark.c
index 90ea380f..ca0c6ec2 100644
--- a/test/benchmark.c
+++ b/test/benchmark.c
@@ -24,8 +24,10 @@
 struct buffers {
 	char *reg;
 	char *enc;
+	uint16_t *enc16;
 	size_t regsz;
 	size_t encsz;
+	bool   enc16ready;
 };
 
 // Define buffer sizes to test with:
@@ -36,11 +38,11 @@ static struct bufsize {
 	int	 batch;
 }
 sizes[] = {
-	{ "10 MB",	MB * 10,	10,	1	},
-	{ "1 MB",	MB * 1,		10,	10	},
-	{ "100 KB",	KB * 100,	10,	100	},
-	{ "10 KB",	KB * 10,	100,	100	},
-	{ "1 KB",	KB * 1,		100,	1000	},
+	{ "10 MB",	MB * 10,	100,	1	},
+	{ "1 MB",	MB * 1,		100,	10	},
+	{ "100 KB",	KB * 100,	100,	100	},
+	{ "10 KB",	KB * 10,	1000,	100	},
+	{ "1 KB",	KB * 1,		1000,	1000	},
 };
 
 static inline float
@@ -134,7 +136,7 @@ codec_bench_enc (struct buffers *b, const struct bufsize *bs, const char *name,
 			fastest = timediff;
 	}
 
-	printf("%s\tencode\t%.02f MB/sec\n", name, bytes_to_mb(b->regsz) / fastest);
+	//printf("%-6s encode   %8.2f MB/sec\n", name, bytes_to_mb(b->regsz) / fastest);
 }
 
 static void
@@ -163,14 +165,56 @@ codec_bench_dec (struct buffers *b, const struct bufsize *bs, const char *name,
 			fastest = timediff;
 	}
 
-	printf("%s\tdecode\t%.02f MB/sec\n", name, bytes_to_mb(b->encsz) / fastest);
+	printf("%-6s decode   %8.2f MB/sec\n", name, bytes_to_mb(b->encsz) / fastest);
+}
+
+static void
+codec_bench_dec16 (struct buffers *b, const struct bufsize *bs, const char *name, unsigned int flags)
+{
+	float timediff, fastest = -1.0f;
+	base64_timespec start, end;
+
+	// Reset buffer size:
+	b->encsz = bs->len;
+
+	// Repeat benchmark a number of times for a fair test:
+	for (int i = bs->repeat; i; i--) {
+
+		// Timing loop, use batches to increase timer resolution:
+		base64_gettime(&start);
+		for (int j = bs->batch; j; j--)
+			base64_decode16(b->enc16, b->encsz, b->reg, &b->regsz, flags);
+		base64_gettime(&end);
+
+		// Calculate average time of batch:
+		timediff = timediff_sec(&start, &end) / bs->batch;
+
+		// Update fastest time seen:
+		if (fastest < 0.0f || timediff < fastest)
+			fastest = timediff;
+	}
+
+	printf("%-6s decode16 %8.2f MB/sec\n", name, bytes_to_mb(b->encsz) / fastest);
+}
+
+static void
+codec_bench_prepare_enc16 (struct buffers *b)
+{
+	for (size_t i = 0U; i < b->encsz; ++i) {
+		b->enc16[i] = (uint8_t)b->enc[i];
+	}
 }
 
 static void
 codec_bench (struct buffers *b, const struct bufsize *bs, const char *name, unsigned int flags)
 {
 	codec_bench_enc(b, bs, name, flags);
+	if (!b->enc16ready) {
+		codec_bench_prepare_enc16(b);
+		b->enc16ready = true;
+	}
 	codec_bench_dec(b, bs, name, flags);
+	codec_bench_dec16(b, bs, name, flags);
 }
 
 int
@@ -183,6 +227,7 @@ main ()
 	// Set buffer sizes to largest buffer length:
 	b.regsz = sizes[0].len;
 	b.encsz = sizes[0].len * 5 / 3;
+	b.enc16ready = false;
 
 	// Allocate space for megabytes of random data:
 	if ((b.reg = malloc(b.regsz)) == NULL) {
@@ -198,10 +243,17 @@ main ()
 		goto err1;
 	}
 
+	// Allocate space for UTF-16 encoded output:
+	if ((b.enc16 = malloc(b.encsz * sizeof(uint16_t))) == NULL) {
+		errmsg = "Out of memory";
+		ret = 1;
+		goto err2;
+	}
+
 	// Fill buffer with random data:
 	if (get_random_data(&b, &errmsg) == false) {
 		ret = 1;
-		goto err2;
+		goto err3;
 	}
 
 	// Loop over all buffer sizes:
@@ -216,6 +268,7 @@ main ()
 	};
 
 	// Free memory:
+err3: free(b.enc16);
 err2:	free(b.enc);
 err1:	free(b.reg);
 err0:	if (errmsg)
diff --git a/test/codec_supported.h b/test/codec_supported.h
index 5ece82f4..18baec08 100644
--- a/test/codec_supported.h
+++ b/test/codec_supported.h
@@ -1,3 +1,3 @@
-char **codecs;
+extern char **codecs;
 
 int codec_supported (int flags);
diff --git a/test/moby_dick.h b/test/moby_dick.h
index 19705355..1d03b6fa 100644
--- a/test/moby_dick.h
+++ b/test/moby_dick.h
@@ -39,3 +39,67 @@ static const char *moby_dick_base64 =
 	"0IGtuZXcKaXQsIGFsbW9zdCBhbGwgbWVuIGluIHRoZWlyIGRlZ3JlZSwgc29tZSB0aW1l"
 	"IG9yIG90aGVyLCBjaGVyaXNoIHZlcnkKbmVhcmx5IHRoZSBzYW1lIGZlZWxpbmdzIHRvd"
 	"2FyZHMgdGhlIG9jZWFuIHdpdGggbWUuCg==";
+
+static const char *moby_dick_base64_16u =
+	"Q\0""2\0""F\0""s\0""b\0""C\0""B\0""t\0""Z\0""S\0""B\0""J\0""c\0""2\0""h\0""t\0""Y\0""W\0""V\0""s\0""L\0""i\0""B\0""T\0"
+	"b\0""2\0""1\0""l\0""I\0""H\0""l\0""l\0""Y\0""X\0""J\0""z\0""I\0""G\0""F\0""n\0""b\0""y\0""0\0""t\0""b\0""m\0""V\0""2\0"
+	"Z\0""X\0""I\0""g\0""b\0""W\0""l\0""u\0""Z\0""C\0""B\0""o\0""b\0""3\0""c\0""g\0""b\0""G\0""9\0""u\0""Z\0""y\0""B\0""w\0"
+	"c\0""m\0""V\0""j\0""a\0""X\0""N\0""l\0""b\0""H\0""k\0""t\0""L\0""W\0""h\0""h\0""d\0""m\0""l\0""u\0""Z\0""w\0""p\0""s\0"
+	"a\0""X\0""R\0""0\0""b\0""G\0""U\0""g\0""b\0""3\0""I\0""g\0""b\0""m\0""8\0""g\0""b\0""W\0""9\0""u\0""Z\0""X\0""k\0""g\0"
+	"a\0""W\0""4\0""g\0""b\0""X\0""k\0""g\0""c\0""H\0""V\0""y\0""c\0""2\0""U\0""s\0""I\0""G\0""F\0""u\0""Z\0""C\0""B\0""u\0"
+	"b\0""3\0""R\0""o\0""a\0""W\0""5\0""n\0""I\0""H\0""B\0""h\0""c\0""n\0""R\0""p\0""Y\0""3\0""V\0""s\0""Y\0""X\0""I\0""g\0"
+	"d\0""G\0""8\0""g\0""a\0""W\0""5\0""0\0""Z\0""X\0""J\0""l\0""c\0""3\0""Q\0""g\0""b\0""W\0""U\0""g\0""b\0""2\0""4\0""K\0"
+	"c\0""2\0""h\0""v\0""c\0""m\0""U\0""s\0""I\0""E\0""k\0""g\0""d\0""G\0""h\0""v\0""d\0""W\0""d\0""o\0""d\0""C\0""B\0""J\0"
+	"I\0""H\0""d\0""v\0""d\0""W\0""x\0""k\0""I\0""H\0""N\0""h\0""a\0""W\0""w\0""g\0""Y\0""W\0""J\0""v\0""d\0""X\0""Q\0""g\0"
+	"Y\0""S\0""B\0""s\0""a\0""X\0""R\0""0\0""b\0""G\0""U\0""g\0""Y\0""W\0""5\0""k\0""I\0""H\0""N\0""l\0""Z\0""S\0""B\0""0\0"
+	"a\0""G\0""U\0""g\0""d\0""2\0""F\0""0\0""Z\0""X\0""J\0""5\0""I\0""H\0""B\0""h\0""c\0""n\0""Q\0""g\0""b\0""2\0""Y\0""K\0"
+	"d\0""G\0""h\0""l\0""I\0""H\0""d\0""v\0""c\0""m\0""x\0""k\0""L\0""i\0""B\0""J\0""d\0""C\0""B\0""p\0""c\0""y\0""B\0""h\0"
+	"I\0""H\0""d\0""h\0""e\0""S\0""B\0""J\0""I\0""G\0""h\0""h\0""d\0""m\0""U\0""g\0""b\0""2\0""Y\0""g\0""Z\0""H\0""J\0""p\0"
+	"d\0""m\0""l\0""u\0""Z\0""y\0""B\0""v\0""Z\0""m\0""Y\0""g\0""d\0""G\0""h\0""l\0""I\0""H\0""N\0""w\0""b\0""G\0""V\0""l\0"
+	"b\0""i\0""B\0""h\0""b\0""m\0""Q\0""g\0""c\0""m\0""V\0""n\0""d\0""W\0""x\0""h\0""d\0""G\0""l\0""u\0""Z\0""w\0""p\0""0\0"
+	"a\0""G\0""U\0""g\0""Y\0""2\0""l\0""y\0""Y\0""3\0""V\0""s\0""Y\0""X\0""R\0""p\0""b\0""2\0""4\0""u\0""I\0""F\0""d\0""o\0"
+	"Z\0""W\0""5\0""l\0""d\0""m\0""V\0""y\0""I\0""E\0""k\0""g\0""Z\0""m\0""l\0""u\0""Z\0""C\0""B\0""t\0""e\0""X\0""N\0""l\0"
+	"b\0""G\0""Y\0""g\0""Z\0""3\0""J\0""v\0""d\0""2\0""l\0""u\0""Z\0""y\0""B\0""n\0""c\0""m\0""l\0""t\0""I\0""G\0""F\0""i\0"
+	"b\0""3\0""V\0""0\0""I\0""H\0""R\0""o\0""Z\0""S\0""B\0""t\0""b\0""3\0""V\0""0\0""a\0""D\0""s\0""K\0""d\0""2\0""h\0""l\0"
+	"b\0""m\0""V\0""2\0""Z\0""X\0""I\0""g\0""a\0""X\0""Q\0""g\0""a\0""X\0""M\0""g\0""Y\0""S\0""B\0""k\0""Y\0""W\0""1\0""w\0"
+	"L\0""C\0""B\0""k\0""c\0""m\0""l\0""6\0""e\0""m\0""x\0""5\0""I\0""E\0""5\0""v\0""d\0""m\0""V\0""t\0""Y\0""m\0""V\0""y\0"
+	"I\0""G\0""l\0""u\0""I\0""G\0""1\0""5\0""I\0""H\0""N\0""v\0""d\0""W\0""w\0""7\0""I\0""H\0""d\0""o\0""Z\0""W\0""5\0""l\0"
+	"d\0""m\0""V\0""y\0""I\0""E\0""k\0""g\0""Z\0""m\0""l\0""u\0""Z\0""A\0""p\0""t\0""e\0""X\0""N\0""l\0""b\0""G\0""Y\0""g\0"
+	"a\0""W\0""5\0""2\0""b\0""2\0""x\0""1\0""b\0""n\0""R\0""h\0""c\0""m\0""l\0""s\0""e\0""S\0""B\0""w\0""Y\0""X\0""V\0""z\0"
+	"a\0""W\0""5\0""n\0""I\0""G\0""J\0""l\0""Z\0""m\0""9\0""y\0""Z\0""S\0""B\0""j\0""b\0""2\0""Z\0""m\0""a\0""W\0""4\0""g\0"
+	"d\0""2\0""F\0""y\0""Z\0""W\0""h\0""v\0""d\0""X\0""N\0""l\0""c\0""y\0""w\0""g\0""Y\0""W\0""5\0""k\0""I\0""G\0""J\0""y\0"
+	"a\0""W\0""5\0""n\0""a\0""W\0""5\0""n\0""I\0""H\0""V\0""w\0""C\0""n\0""R\0""o\0""Z\0""S\0""B\0""y\0""Z\0""W\0""F\0""y\0"
+	"I\0""G\0""9\0""m\0""I\0""G\0""V\0""2\0""Z\0""X\0""J\0""5\0""I\0""G\0""Z\0""1\0""b\0""m\0""V\0""y\0""Y\0""W\0""w\0""g\0"
+	"S\0""S\0""B\0""t\0""Z\0""W\0""V\0""0\0""O\0""y\0""B\0""h\0""b\0""m\0""Q\0""g\0""Z\0""X\0""N\0""w\0""Z\0""W\0""N\0""p\0"
+	"Y\0""W\0""x\0""s\0""e\0""S\0""B\0""3\0""a\0""G\0""V\0""u\0""Z\0""X\0""Z\0""l\0""c\0""i\0""B\0""t\0""e\0""S\0""B\0""o\0"
+	"e\0""X\0""B\0""v\0""c\0""y\0""B\0""n\0""Z\0""X\0""Q\0""K\0""c\0""3\0""V\0""j\0""a\0""C\0""B\0""h\0""b\0""i\0""B\0""1\0"
+	"c\0""H\0""B\0""l\0""c\0""i\0""B\0""o\0""Y\0""W\0""5\0""k\0""I\0""G\0""9\0""m\0""I\0""G\0""1\0""l\0""L\0""C\0""B\0""0\0"
+	"a\0""G\0""F\0""0\0""I\0""G\0""l\0""0\0""I\0""H\0""J\0""l\0""c\0""X\0""V\0""p\0""c\0""m\0""V\0""z\0""I\0""G\0""E\0""g\0"
+	"c\0""3\0""R\0""y\0""b\0""2\0""5\0""n\0""I\0""G\0""1\0""v\0""c\0""m\0""F\0""s\0""I\0""H\0""B\0""y\0""a\0""W\0""5\0""j\0"
+	"a\0""X\0""B\0""s\0""Z\0""S\0""B\0""0\0""b\0""w\0""p\0""w\0""c\0""m\0""V\0""2\0""Z\0""W\0""5\0""0\0""I\0""G\0""1\0""l\0"
+	"I\0""G\0""Z\0""y\0""b\0""2\0""0\0""g\0""Z\0""G\0""V\0""s\0""a\0""W\0""J\0""l\0""c\0""m\0""F\0""0\0""Z\0""W\0""x\0""5\0"
+	"I\0""H\0""N\0""0\0""Z\0""X\0""B\0""w\0""a\0""W\0""5\0""n\0""I\0""G\0""l\0""u\0""d\0""G\0""8\0""g\0""d\0""G\0""h\0""l\0"
+	"I\0""H\0""N\0""0\0""c\0""m\0""V\0""l\0""d\0""C\0""w\0""g\0""Y\0""W\0""5\0""k\0""I\0""G\0""1\0""l\0""d\0""G\0""h\0""v\0"
+	"Z\0""G\0""l\0""j\0""Y\0""W\0""x\0""s\0""e\0""Q\0""p\0""r\0""b\0""m\0""9\0""j\0""a\0""2\0""l\0""u\0""Z\0""y\0""B\0""w\0"
+	"Z\0""W\0""9\0""w\0""b\0""G\0""U\0""n\0""c\0""y\0""B\0""o\0""Y\0""X\0""R\0""z\0""I\0""G\0""9\0""m\0""Z\0""i\0""0\0""t\0"
+	"d\0""G\0""h\0""l\0""b\0""i\0""w\0""g\0""S\0""S\0""B\0""h\0""Y\0""2\0""N\0""v\0""d\0""W\0""5\0""0\0""I\0""G\0""l\0""0\0"
+	"I\0""G\0""h\0""p\0""Z\0""2\0""g\0""g\0""d\0""G\0""l\0""t\0""Z\0""S\0""B\0""0\0""b\0""y\0""B\0""n\0""Z\0""X\0""Q\0""g\0"
+	"d\0""G\0""8\0""g\0""c\0""2\0""V\0""h\0""C\0""m\0""F\0""z\0""I\0""H\0""N\0""v\0""b\0""2\0""4\0""g\0""Y\0""X\0""M\0""g\0"
+	"S\0""S\0""B\0""j\0""Y\0""W\0""4\0""u\0""I\0""F\0""R\0""o\0""a\0""X\0""M\0""g\0""a\0""X\0""M\0""g\0""b\0""X\0""k\0""g\0"
+	"c\0""3\0""V\0""i\0""c\0""3\0""R\0""p\0""d\0""H\0""V\0""0\0""Z\0""S\0""B\0""m\0""b\0""3\0""I\0""g\0""c\0""G\0""l\0""z\0"
+	"d\0""G\0""9\0""s\0""I\0""G\0""F\0""u\0""Z\0""C\0""B\0""i\0""Y\0""W\0""x\0""s\0""L\0""i\0""B\0""X\0""a\0""X\0""R\0""o\0"
+	"I\0""G\0""E\0""K\0""c\0""G\0""h\0""p\0""b\0""G\0""9\0""z\0""b\0""3\0""B\0""o\0""a\0""W\0""N\0""h\0""b\0""C\0""B\0""m\0"
+	"b\0""G\0""9\0""1\0""c\0""m\0""l\0""z\0""a\0""C\0""B\0""D\0""Y\0""X\0""R\0""v\0""I\0""H\0""R\0""o\0""c\0""m\0""9\0""3\0"
+	"c\0""y\0""B\0""o\0""a\0""W\0""1\0""z\0""Z\0""W\0""x\0""m\0""I\0""H\0""V\0""w\0""b\0""2\0""4\0""g\0""a\0""G\0""l\0""z\0"
+	"I\0""H\0""N\0""3\0""b\0""3\0""J\0""k\0""O\0""y\0""B\0""J\0""I\0""H\0""F\0""1\0""a\0""W\0""V\0""0\0""b\0""H\0""k\0""K\0"
+	"d\0""G\0""F\0""r\0""Z\0""S\0""B\0""0\0""b\0""y\0""B\0""0\0""a\0""G\0""U\0""g\0""c\0""2\0""h\0""p\0""c\0""C\0""4\0""g\0"
+	"V\0""G\0""h\0""l\0""c\0""m\0""U\0""g\0""a\0""X\0""M\0""g\0""b\0""m\0""9\0""0\0""a\0""G\0""l\0""u\0""Z\0""y\0""B\0""z\0"
+	"d\0""X\0""J\0""w\0""c\0""m\0""l\0""z\0""a\0""W\0""5\0""n\0""I\0""G\0""l\0""u\0""I\0""H\0""R\0""o\0""a\0""X\0""M\0""u\0"
+	"I\0""E\0""l\0""m\0""I\0""H\0""R\0""o\0""Z\0""X\0""k\0""g\0""Y\0""n\0""V\0""0\0""I\0""G\0""t\0""u\0""Z\0""X\0""c\0""K\0"
+	"a\0""X\0""Q\0""s\0""I\0""G\0""F\0""s\0""b\0""W\0""9\0""z\0""d\0""C\0""B\0""h\0""b\0""G\0""w\0""g\0""b\0""W\0""V\0""u\0"
+	"I\0""G\0""l\0""u\0""I\0""H\0""R\0""o\0""Z\0""W\0""l\0""y\0""I\0""G\0""R\0""l\0""Z\0""3\0""J\0""l\0""Z\0""S\0""w\0""g\0"
+	"c\0""2\0""9\0""t\0""Z\0""S\0""B\0""0\0""a\0""W\0""1\0""l\0""I\0""G\0""9\0""y\0""I\0""G\0""9\0""0\0""a\0""G\0""V\0""y\0"
+	"L\0""C\0""B\0""j\0""a\0""G\0""V\0""y\0""a\0""X\0""N\0""o\0""I\0""H\0""Z\0""l\0""c\0""n\0""k\0""K\0""b\0""m\0""V\0""h\0"
+	"c\0""m\0""x\0""5\0""I\0""H\0""R\0""o\0""Z\0""S\0""B\0""z\0""Y\0""W\0""1\0""l\0""I\0""G\0""Z\0""l\0""Z\0""W\0""x\0""p\0"
+	"b\0""m\0""d\0""z\0""I\0""H\0""R\0""v\0""d\0""2\0""F\0""y\0""Z\0""H\0""M\0""g\0""d\0""G\0""h\0""l\0""I\0""G\0""9\0""j\0"
+	"Z\0""W\0""F\0""u\0""I\0""H\0""d\0""p\0""d\0""G\0""g\0""g\0""b\0""W\0""U\0""u\0""C\0""g\0""=\0""=\0";
diff --git a/test/test_base64.c b/test/test_base64.c
index 15996914..1c87b420 100644
--- a/test/test_base64.c
+++ b/test/test_base64.c
@@ -57,6 +57,32 @@ assert_dec (int flags, const char *src, const char *dst)
 	return false;
 }
 
+static bool
+assert_dec16 (int flags, const char *src, const char *src16, const char *dst)
+{
+	size_t srclen = strlen(src);
+	size_t dstlen = strlen(dst);
+
+	if (!base64_decode16((const uint16_t*)src16, srclen, out, &outlen, flags)) {
+		printf("FAIL: decoding of '%s': decoding error\n", src);
+		return true;
+	}
+	if (outlen != dstlen) {
+		printf("FAIL: encoding of '%s': "
+					 "length expected %lu, got %lu\n", src,
+					 (unsigned long)dstlen,
+					 (unsigned long)outlen
+					 );
+		return true;
+	}
+	if (strncmp(dst, out, outlen) != 0) {
+		out[outlen] = '\0';
+		printf("FAIL: decoding of '%s': expected output '%s', got '%s'\n", src, dst, out);
+		return true;
+	}
+	return false;
+}
+
 static int
 assert_roundtrip (int flags, const char *src)
 {
@@ -292,22 +318,50 @@ test_one_codec (const char *codec, int flags)
 	struct {
 		const char *in;
 		const char *out;
+		const char *out16;
 	} vec[] = {
 
 		// These are the test vectors from RFC4648:
-		{ "",		""         },
-		{ "f",		"Zg=="     },
-		{ "fo",		"Zm8="     },
-		{ "foo",	"Zm9v"     },
-		{ "foob",	"Zm9vYg==" },
-		{ "fooba",	"Zm9vYmE=" },
-		{ "foobar",	"Zm9vYmFy" },
+		{ "",		""        , "\0" },
+		{ "f",		"Zg=="    , "Z\0g\0=\0=\0" },
+		{ "fo",		"Zm8="    , "Z\0m\0""8\0=\0" },
+		{ "foo",	"Zm9v"    , "Z\0m\0""9\0v\0" },
+		{ "foob",	"Zm9vYg==", "Z\0m\0""9\0v\0Y\0g\0=\0=\0" },
+		{ "fooba",	"Zm9vYmE=", "Z\0m\0""9\0v\0Y\0m\0E\0=\0" },
+		{ "foobar",	"Zm9vYmFy", "Z\0m\0""9\0v\0Y\0m\0F\0y\0" },
 
 		// The first paragraph from Moby Dick,
 		// to test the SIMD codecs with larger blocksize:
-		{ moby_dick_plain, moby_dick_base64 },
+		{ moby_dick_plain, moby_dick_base64, moby_dick_base64_16u },
 	};
 
+#if 0
+	size_t c = 0;
+	const char * pSrc = moby_dick_base64;
+	putc('\t', stdout);
+	putc('\"', stdout);
+	while (*pSrc) {
+		int character = *pSrc++;
+		putc(character, stdout);
+		putc('\\', stdout);
+		putc('0', stdout);
+		putc('\"', stdout);
+		putc('\"', stdout);
+
+		c++;
+		if (c == 24U) {
+			putc('\"', stdout);
+			putc('\n', stdout);
+			putc('\t', stdout);
+			putc('\"', stdout);
+
+			c = 0U;
+		}
+	}
+	putc('\"', stdout);
+	putc('\n', stdout);
+#endif
+
 	for (size_t i = 0; i < sizeof(vec) / sizeof(vec[0]); i++) {
 
 		// Encode plain string, check against output:
@@ -315,6 +369,7 @@ test_one_codec (const char *codec, int flags)
 
 		// Decode the output string, check if we get the input:
 		fail |= assert_dec(flags, vec[i].out, vec[i].in);
+		fail |= assert_dec16(flags, vec[i].out, vec[i].out16, vec[i].in);
 
 		// Do a roundtrip on the inputs and the outputs:
 		fail |= assert_roundtrip(flags, vec[i].in);