diff --git a/libfreerdp/primitives/test/CMakeLists.txt b/libfreerdp/primitives/test/CMakeLists.txt
index 738329152..6d6898f17 100644
--- a/libfreerdp/primitives/test/CMakeLists.txt
+++ b/libfreerdp/primitives/test/CMakeLists.txt
@@ -31,11 +31,28 @@ set(PRIMITIVE_TEST_CFILES
 	test_set.c
 	test_shift.c
 	test_sign.c
+	../prim_add.c
+	../prim_andor.c
+	../prim_alphaComp.c
+	../prim_colors.c
+	../prim_copy.c
+	../prim_set.c
+	../prim_shift.c
+	../prim_sign.c
+	../prim_add_opt.c
+	../prim_alphaComp_opt.c
+	../prim_andor_opt.c
+	../prim_colors_opt.c
+	../prim_set_opt.c
+	../prim_shift_opt.c
+	../prim_sign_opt.c
+	../primitives.c
     )
 
 set(PRIMITIVE_TEST_HEADERS
 	measure.h
 	prim_test.h
+	../prim_internal.h
 )
 
 set(PRIMITIVE_TEST_SRCS
@@ -121,7 +138,7 @@ endif()
 
 set_property(SOURCE ${PRIMITIVE_TEST_CFILES} PROPERTY COMPILE_FLAGS ${OPTFLAGS})
 
-target_link_libraries(prim_test freerdp-primitives rt winpr-sysinfo)
+target_link_libraries(prim_test rt winpr-sysinfo)
 if(NOT TESTING_OUTPUT_DIRECTORY)
 	set(TESTING_OUTPUT_DIRECTORY .)
 endif()
diff --git a/libfreerdp/primitives/test/prim_test.c b/libfreerdp/primitives/test/prim_test.c
index 7e88efbc9..5104eebc7 100644
--- a/libfreerdp/primitives/test/prim_test.c
+++ b/libfreerdp/primitives/test/prim_test.c
@@ -44,8 +44,8 @@ typedef struct
 } flagpair_t;
 
 static const flagpair_t flags[] =
-#ifdef _M_IX86_AMD64
 {
+#ifdef _M_IX86_AMD64
 	{ PF_MMX_INSTRUCTIONS_AVAILABLE,		"MMX" },
 	{ PF_3DNOW_INSTRUCTIONS_AVAILABLE,		"3DNow" },
 	{ PF_XMMI_INSTRUCTIONS_AVAILABLE,		"SSE" },
diff --git a/libfreerdp/primitives/test/prim_test.h b/libfreerdp/primitives/test/prim_test.h
index fa61025f0..f32f46889 100644
--- a/libfreerdp/primitives/test/prim_test.h
+++ b/libfreerdp/primitives/test/prim_test.h
@@ -101,7 +101,7 @@ extern int test_or_32u_speed(void);
 /* Since so much of this code is repeated, define a macro to build 
  * functions to do speed tests.
  */
-#ifdef armel
+#ifdef _M_ARM
 #define SIMD_TYPE "Neon"
 #else
 #define SIMD_TYPE "SSE"
@@ -122,7 +122,7 @@ extern int test_or_32u_speed(void);
 		} \
 	} while (0)
 
-#if (defined(_M_IX86_AMD64) && defined(WITH_SSE2)) || (defined(arm) && defined(WITH_NEON))
+#if (defined(_M_IX86_AMD64) && defined(WITH_SSE2)) || (defined(_M_ARM) && defined(WITH_NEON))
 #define DO_OPT_MEASUREMENTS(_funcOpt_, _prework_) \
 	do { \
 		for (s=0; s<num_sizes; ++s) \
@@ -160,6 +160,7 @@ extern int test_or_32u_speed(void);
 #define DO_IPP_MEASUREMENTS(_funcIPP_, _prework_)
 #endif
 
+#define PRIM_NOP do {} while (0)
 /* ------------------------------------------------------------------------- */
 #define STD_SPEED_TEST( \
 	_name_, _srctype_, _dsttype_, _prework_, \
diff --git a/libfreerdp/primitives/test/test_add.c b/libfreerdp/primitives/test/test_add.c
index 083c37907..453e2a7de 100644
--- a/libfreerdp/primitives/test/test_add.c
+++ b/libfreerdp/primitives/test/test_add.c
@@ -34,7 +34,9 @@ int test_add16s_func(void)
 	INT16 ALIGN(src1[FUNC_TEST_SIZE+3]), ALIGN(src2[FUNC_TEST_SIZE+3]), 
 		ALIGN(d1[FUNC_TEST_SIZE+3]), ALIGN(d2[FUNC_TEST_SIZE+3]);
 	int failed = 0;
+#if defined(WITH_SSE2) || defined(WITH_IPP)
 	int i;
+#endif
 	char testStr[256];
 
 	testStr[0] = '\0';
@@ -43,7 +45,7 @@ int test_add16s_func(void)
 	memset(d1, 0, sizeof(d1));
 	memset(d2, 0, sizeof(d2));
 	general_add_16s(src1+1, src2+1, d1+1, FUNC_TEST_SIZE);
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 	if(IsProcessorFeaturePresent(PF_SSE3_INSTRUCTIONS_AVAILABLE))
 	{
 		strcat(testStr, " SSE3");
@@ -70,7 +72,7 @@ int test_add16s_func(void)
 			}
 		}
 	}
-#endif /* i386 */
+#endif
 #ifdef WITH_IPP
 	strcat(testStr, " IPP");
 	ippsAdd_16s(src1+1, src2+1, d2+1, FUNC_TEST_SIZE);
@@ -91,7 +93,11 @@ int test_add16s_func(void)
 /* ------------------------------------------------------------------------- */
 STD_SPEED_TEST(add16s_speed_test, INT16, INT16, dst=dst,
 	TRUE, general_add_16s(src1, src2, dst, size),
+#ifdef WITH_SSE2
 	TRUE, sse3_add_16s(src1, src2, dst, size), PF_SSE3_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
 	TRUE, ippsAdd_16s(src1, src2, dst, size));
 
 int test_add16s_speed(void)
diff --git a/libfreerdp/primitives/test/test_alphaComp.c b/libfreerdp/primitives/test/test_alphaComp.c
index 2d8285f27..7d721cd39 100644
--- a/libfreerdp/primitives/test/test_alphaComp.c
+++ b/libfreerdp/primitives/test/test_alphaComp.c
@@ -132,7 +132,7 @@ int test_alphaComp_func(void)
 	general_alphaComp_argb((const BYTE *) src1, 4*SRC1_WIDTH, 
 		(const BYTE *) src2, 4*SRC2_WIDTH,
 		(BYTE *) dst1, 4*DST_WIDTH, TEST_WIDTH, TEST_HEIGHT);
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 	if (IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
 	{
 		strcat(testStr, " SSE2");
@@ -165,7 +165,7 @@ int test_alphaComp_func(void)
 					x, y, s1, s2, c0, c1);
 				error = 1;
 			}
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 			if (IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
 			{
 				UINT32 c2 = *PIXEL(dst2a, 4*DST_WIDTH, x, y);
@@ -206,8 +206,13 @@ int test_alphaComp_func(void)
 STD_SPEED_TEST(alphaComp_speed, BYTE, BYTE, int bytes = size*4,
 	TRUE, general_alphaComp_argb(src1, bytes, src2, bytes, dst, bytes,
 		size, size),
+#ifdef WITH_SSE2
 	TRUE, sse2_alphaComp_argb(src1, bytes, src2, bytes, dst, bytes,
 		size, size), PF_XMMI64_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
+
 	TRUE, ipp_alphaComp_argb(src1, bytes, src2, bytes, dst, bytes,
 		size, size));
 
diff --git a/libfreerdp/primitives/test/test_andor.c b/libfreerdp/primitives/test/test_andor.c
index 6e8b3d8a3..d82991408 100644
--- a/libfreerdp/primitives/test/test_andor.c
+++ b/libfreerdp/primitives/test/test_andor.c
@@ -55,7 +55,7 @@ int test_and_32u_func(void)
 			++failed;
 		}
 	}
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 	if (IsProcessorFeaturePresent(PF_SSE3_INSTRUCTIONS_AVAILABLE))
 	{
 		strcat(testStr, " SSE3");
@@ -92,7 +92,11 @@ int test_and_32u_func(void)
 /* ------------------------------------------------------------------------- */
 STD_SPEED_TEST(andC_32u_speed_test, UINT32, UINT32, dst=dst,
 	TRUE, general_andC_32u(src1, constant, dst, size),
+#ifdef WITH_SSE2
 	TRUE, sse3_andC_32u(src1, constant, dst, size), PF_SSE3_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
 	TRUE, ippsAndC_32u(src1, constant, dst, size))
 
 int test_and_32u_speed(void)
@@ -127,7 +131,7 @@ int test_or_32u_func(void)
 		++failed;
 		}
 	}
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 	if(IsProcessorFeaturePresent(PF_SSE3_INSTRUCTIONS_AVAILABLE))
 	{
 		strcat(testStr, " SSE3");
@@ -164,7 +168,11 @@ int test_or_32u_func(void)
 /* ------------------------------------------------------------------------- */
 STD_SPEED_TEST(orC_32u_speed_test, UINT32, UINT32, dst=dst,
 	TRUE, general_orC_32u(src1, constant, dst, size),
+#ifdef WITH_SSE2
 	TRUE, sse3_orC_32u(src1, constant, dst, size), PF_SSE3_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
 	TRUE, ippsOrC_32u(src1, constant, dst, size))
 
 int test_or_32u_speed(void)
diff --git a/libfreerdp/primitives/test/test_colors.c b/libfreerdp/primitives/test/test_colors.c
index e5192c035..57e568b9c 100644
--- a/libfreerdp/primitives/test/test_colors.c
+++ b/libfreerdp/primitives/test/test_colors.c
@@ -31,12 +31,17 @@ extern pstatus_t general_yCbCrToRGB_16s16s_P3P3(const INT16 *pSrc[3],
 	int srcStep, INT16 *pDst[3], int dstStep, const prim_size_t *roi);
 extern pstatus_t sse2_yCbCrToRGB_16s16s_P3P3(const INT16 *pSrc[3],
 	int srcStep, INT16 *pDst[3], int dstStep, const prim_size_t *roi);
+extern pstatus_t neon_yCbCrToRGB_16s16s_P3P3(const INT16 *pSrc[3],
+	int srcStep, INT16 *pDst[3], int dstStep, const prim_size_t *roi);
 
 /* ------------------------------------------------------------------------- */
 int test_RGBToRGB_16s8u_P3AC4R_func(void)
 {
 	INT16 ALIGN(r[4096]), ALIGN(g[4096]), ALIGN(b[4096]);
-	UINT32 ALIGN(out1[4096]), ALIGN(out2[4096]);
+	UINT32 ALIGN(out1[4096]);
+#ifdef WITH_SSE2
+	UINT32 ALIGN(out2[4096]);
+#endif
 	int i;
 	int failed = 0;
 	char testStr[256];
@@ -61,7 +66,7 @@ int test_RGBToRGB_16s8u_P3AC4R_func(void)
 
 	general_RGBToRGB_16s8u_P3AC4R((const INT16 **) ptrs, 64*2,
 		(BYTE *) out1, 64*4, &roi);
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 	if (IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
 	{
 		strcat(testStr, " SSE2");
@@ -88,9 +93,13 @@ STD_SPEED_TEST(
 	rgb_to_argb_speed, INT16*, UINT32, dst=dst,
 	TRUE, general_RGBToRGB_16s8u_P3AC4R(
 		(const INT16 **) src1, 64*2, (BYTE *) dst, 64*4, &roi64x64),
+#ifdef WITH_SSE2
 	TRUE, sse2_RGBToRGB_16s8u_P3AC4R(
 		(const INT16 **) src1, 64*2, (BYTE *) dst, 64*4, &roi64x64),
 		PF_XMMI64_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
 	FALSE, dst=dst);
 
 int test_RGBToRGB_16s8u_P3AC4R_speed(void)
@@ -165,7 +174,7 @@ int test_yCbCrToRGB_16s16s_P3P3_func(void)
 	out2[2] = b2;
 
 	general_yCbCrToRGB_16s16s_P3P3(in, 64*2, out1, 64*2, &roi);
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 	if (IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
 	{
 		strcat(testStr, " SSE2");
@@ -190,8 +199,15 @@ int test_yCbCrToRGB_16s16s_P3P3_func(void)
 STD_SPEED_TEST(
 	ycbcr_to_rgb_speed, INT16*, INT16*, dst=dst,
 	TRUE, general_yCbCrToRGB_16s16s_P3P3(src1, 64*2, dst, 64*2, &roi64x64),
+#ifdef WITH_SSE2
 	TRUE, sse2_yCbCrToRGB_16s16s_P3P3(src1, 64*2, dst, 64*2, &roi64x64),
 		PF_XMMI64_INSTRUCTIONS_AVAILABLE, FALSE,
+#elif defined(WITH_NEON)
+	TRUE, neon_yCbCrToRGB_16s16s_P3P3(src1, 64*2, dst, 64*2, &roi64x64),
+		PF_ARM_NEON_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
 	FALSE, dst=dst);
 
 int test_yCbCrToRGB_16s16s_P3P3_speed(void)
diff --git a/libfreerdp/primitives/test/test_copy.c b/libfreerdp/primitives/test/test_copy.c
index 5f2ed9103..13b4f57f2 100644
--- a/libfreerdp/primitives/test/test_copy.c
+++ b/libfreerdp/primitives/test/test_copy.c
@@ -71,7 +71,7 @@ int test_copy8u_func(void)
 /* ------------------------------------------------------------------------- */
 STD_SPEED_TEST(copy8u_speed_test, BYTE, BYTE, dst=dst,
 	TRUE, memcpy(dst, src1, size),
-	FALSE, NULL, 0, FALSE,
+	FALSE, PRIM_NOP, 0, FALSE,
 	TRUE, ippsCopy_8u(src1, dst, size));
 
 int test_copy8u_speed(void)
diff --git a/libfreerdp/primitives/test/test_set.c b/libfreerdp/primitives/test/test_set.c
index 0343674f5..33a41e216 100644
--- a/libfreerdp/primitives/test/test_set.c
+++ b/libfreerdp/primitives/test/test_set.c
@@ -37,13 +37,15 @@ static const int set_sizes[] = { 1, 4, 16, 32, 64, 256, 1024, 4096 };
 /* ------------------------------------------------------------------------- */
 int test_set8u_func(void)
 {
+#if defined(WITH_SSE2) || defined(WITH_IPP)
 	BYTE ALIGN(dest[48]);
-	int failed = 0;
 	int off;
+#endif
+	int failed = 0;
 	char testStr[256];
 	testStr[0] = '\0';
 
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 	/* Test SSE under various alignments */
 	if (IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
 	{
@@ -101,7 +103,7 @@ int test_set8u_func(void)
 /* ------------------------------------------------------------------------- */
 STD_SPEED_TEST(set8u_speed_test, BYTE, BYTE, dst=dst,
 	TRUE, memset(dst, constant, size),
-	FALSE, NULL, 0, FALSE,
+	FALSE, PRIM_NOP, 0, FALSE,
 	TRUE, ippsSet_8u(constant, dst, size));
 
 int test_set8u_speed(void)
@@ -115,13 +117,15 @@ int test_set8u_speed(void)
 /* ------------------------------------------------------------------------- */
 int test_set32s_func(void)
 {
+#if defined(WITH_SSE2) || defined(WITH_IPP)
 	INT32 ALIGN(dest[512]);
-	int failed = 0;
 	int off;
+#endif
+	int failed = 0;
 	char testStr[256];
 	testStr[0] = '\0';
 
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 	/* Test SSE under various alignments */
 	if (IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
 	{
@@ -176,13 +180,15 @@ int test_set32s_func(void)
 /* ------------------------------------------------------------------------- */
 int test_set32u_func(void)
 {
+#if defined(WITH_SSE2) || defined(WITH_IPP)
 	UINT32 ALIGN(dest[512]);
-	int failed = 0;
 	int off;
+#endif
+	int failed = 0;
 	char testStr[256];
 	testStr[0] = '\0';
 
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 	/* Test SSE under various alignments */
 	if (IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
 	{
@@ -246,7 +252,11 @@ static inline void memset32u_naive(
 /* ------------------------------------------------------------------------- */
 STD_SPEED_TEST(set32u_speed_test, UINT32, UINT32, dst=dst,
 	TRUE, memset32u_naive(constant, dst, size),
+#ifdef WITH_SSE2
 	TRUE, sse2_set_32u(constant, dst, size), PF_XMMI64_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
 	TRUE, ipp_wrapper_set_32u(constant, dst, size));
 
 int test_set32u_speed(void)
@@ -274,7 +284,11 @@ static inline void memset32s_naive(
 /* ------------------------------------------------------------------------- */
 STD_SPEED_TEST(set32s_speed_test, INT32, INT32, dst=dst,
 	TRUE, memset32s_naive(constant, dst, size),
+#ifdef WITH_SSE2
 	TRUE, sse2_set_32s(constant, dst, size), PF_XMMI64_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, NULL, 0, FALSE,
+#endif
 	TRUE, ippsSet_32s(constant, dst, size));
 
 int test_set32s_speed(void)
diff --git a/libfreerdp/primitives/test/test_shift.c b/libfreerdp/primitives/test/test_shift.c
index d72407e13..2150373d5 100644
--- a/libfreerdp/primitives/test/test_shift.c
+++ b/libfreerdp/primitives/test/test_shift.c
@@ -48,7 +48,7 @@ extern pstatus_t sse2_rShiftC_16u(
 extern pstatus_t sse2_shiftC_16u(
 	const UINT16 *pSrc, int val, UINT16 *pDst, int len);
 
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 #define SHIFT_TEST_FUNC(_name_, _type_, _str_, _f1_, _f2_) \
 int _name_(void) \
 { \
@@ -109,19 +109,35 @@ SHIFT_TEST_FUNC(test_rShift_16u_func, UINT16, "rshift_16u", general_rShiftC_16u,
 /* ========================================================================= */
 STD_SPEED_TEST(speed_lShift_16s, INT16, INT16, dst=dst,
     TRUE, general_lShiftC_16s(src1, constant, dst, size),
+#ifdef WITH_SSE2
 	TRUE, sse2_lShiftC_16s(src1, constant, dst, size), PF_XMMI64_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
 	TRUE, ippsLShiftC_16s(src1, constant, dst, size));
 STD_SPEED_TEST(speed_lShift_16u, UINT16, UINT16, dst=dst,
     TRUE, general_lShiftC_16u(src1, constant, dst, size),
+#ifdef WITH_SSE2
 	TRUE, sse2_lShiftC_16u(src1, constant, dst, size), PF_XMMI64_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
 	TRUE, ippsLShiftC_16u(src1, constant, dst, size));
 STD_SPEED_TEST(speed_rShift_16s, INT16, INT16, dst=dst,
     TRUE, general_rShiftC_16s(src1, constant, dst, size),
+#ifdef WITH_SSE2
 	TRUE, sse2_rShiftC_16s(src1, constant, dst, size), PF_XMMI64_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
 	TRUE, ippsRShiftC_16s(src1, constant, dst, size));
 STD_SPEED_TEST(speed_rShift_16u, UINT16, UINT16, dst=dst,
     TRUE, general_rShiftC_16u(src1, constant, dst, size),
+#ifdef WITH_SSE2
 	TRUE, sse2_rShiftC_16u(src1, constant, dst, size), PF_XMMI64_INSTRUCTIONS_AVAILABLE, FALSE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
 	TRUE, ippsRShiftC_16u(src1, constant, dst, size));
 
 /* ------------------------------------------------------------------------- */
diff --git a/libfreerdp/primitives/test/test_sign.c b/libfreerdp/primitives/test/test_sign.c
index 99f5a60d4..beb22ee1d 100644
--- a/libfreerdp/primitives/test/test_sign.c
+++ b/libfreerdp/primitives/test/test_sign.c
@@ -23,21 +23,26 @@ static const int SIGN_PRETEST_ITERATIONS = 100000;
 static const float TEST_TIME = 1.0;
 
 extern pstatus_t general_sign_16s(const INT16 *pSrc, INT16 *pDst, int len);
+#ifdef WITH_SSE2
 extern pstatus_t ssse3_sign_16s(const INT16 *pSrc, INT16 *pDst, int len);
+#endif
 
 /* ------------------------------------------------------------------------- */
 int test_sign16s_func(void)
 {
-	INT16 ALIGN(src[65535]), ALIGN(d1[65535]), ALIGN(d2[65535]);
-	int failed = 0;
+	INT16 ALIGN(src[65535]), ALIGN(d1[65535]);
+#ifdef WITH_SSE2
+	INT16 ALIGN(d2[65535]);
 	int i;
+#endif
+	int failed = 0;
 	char testStr[256];
 
 	/* Test when we can reach 16-byte alignment */
 	testStr[0] = '\0';
 	get_random_data(src, sizeof(src));
 	general_sign_16s(src+1, d1+1, 65535);
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 	if (IsProcessorFeaturePresentEx(PF_EX_SSSE3))
 	{
 		strcat(testStr, " SSSE3");
@@ -57,7 +62,7 @@ int test_sign16s_func(void)
 	/* Test when we cannot reach 16-byte alignment */
 	get_random_data(src, sizeof(src));
 	general_sign_16s(src+1, d1+2, 65535);
-#ifdef _M_IX86_AMD64
+#ifdef WITH_SSE2
 	if (IsProcessorFeaturePresentEx(PF_EX_SSSE3))
 	{
 		ssse3_sign_16s(src+1, d2+2, 65535);
@@ -79,7 +84,11 @@ int test_sign16s_func(void)
 /* ------------------------------------------------------------------------- */
 STD_SPEED_TEST(sign16s_speed_test, INT16, INT16, dst=dst,
 	TRUE, general_sign_16s(src1, dst, size),
+#ifdef WITH_SSE2
 	TRUE, ssse3_sign_16s(src1, dst, size), PF_EX_SSSE3, TRUE,
+#else
+	FALSE, PRIM_NOP, 0, FALSE,
+#endif
 	FALSE, dst=dst);
 
 int test_sign16s_speed(void)