diff --git a/kernel/x86/KERNEL.NEHALEM b/kernel/x86/KERNEL.NEHALEM
index 65b03ae50..835520efb 100644
--- a/kernel/x86/KERNEL.NEHALEM
+++ b/kernel/x86/KERNEL.NEHALEM
@@ -1 +1,3 @@
 include $(KERNELDIR)/KERNEL.PENRYN
+SSWAPKERNEL  = ../arm/swap.c
+DSWAPKERNEL  = ../arm/swap.c
diff --git a/utest/CMakeLists.txt b/utest/CMakeLists.txt
index 75e6fad7f..1b426afe7 100644
--- a/utest/CMakeLists.txt
+++ b/utest/CMakeLists.txt
@@ -8,9 +8,32 @@ else ()
     utest_main.c
     test_amax.c
     test_rotmg.c
+    test_rot.c
+    test_axpy.c
+    test_dsdot.c
+    test_swap.c
   )
 endif ()
 
+# crashing on travis cl with an error code suggesting resource not found
+if (NOT MSVC)
+set(OpenBLAS_utest_src
+  ${OpenBLAS_utest_src}
+  test_dotu.c
+  )
+endif ()
+
+# known to hang with the native Windows and Android threads
+# FIXME needs checking if this works on any of the other platforms
+if (NOT USE_OPENMP)
+if (OS_CYGWIN_NT OR OS_LINUX)
+set(OpenBLAS_utest_src
+  ${OpenBLAS_utest_src}
+  test_fork.c
+  )
+endif()
+endif()
+
 if (NOT NO_LAPACK)
 set(OpenBLAS_utest_src
   ${OpenBLAS_utest_src}
diff --git a/utest/Makefile b/utest/Makefile
index 625a3bcf4..e40b3c6db 100644
--- a/utest/Makefile
+++ b/utest/Makefile
@@ -8,20 +8,18 @@ UTESTBIN=openblas_utest
 
 include $(TOPDIR)/Makefile.system
 
-OBJS=utest_main.o test_amax.o test_rotmg.o
+OBJS=utest_main.o test_amax.o test_rotmg.o test_axpy.o test_dotu.o test_dsdot.o test_swap.o test_rot.o
 #test_rot.o test_swap.o test_axpy.o test_dotu.o test_dsdot.o test_fork.o
 
 ifneq ($(NO_LAPACK), 1)
-#OBJS += test_potrs.o
+OBJS += test_potrs.o
 endif
 
+#this does not work with OpenMP nor with native Windows or Android threads
+# FIXME TBD if this works on OSX, SunOS, POWER and zarch
 ifndef USE_OPENMP
-ifndef OS_WINDOWS
+ifeq ($(OSNAME), $(filter $(OSNAME),Linux CYGWIN_NT))
 OBJS += test_fork.o
-else
-ifdef OS_CYGWIN_NT
-OBJS += test_fork.o
-endif
 endif
 endif
 
diff --git a/utest/test_axpy.c b/utest/test_axpy.c
index 696eb7a51..783755333 100644
--- a/utest/test_axpy.c
+++ b/utest/test_axpy.c
@@ -31,88 +31,81 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 **********************************************************************************/
 
-#include "common_utest.h"
+#include "openblas_utest.h"
 
-void test_daxpy_inc_0(void)
+CTEST(axpy,daxpy_inc_0)
 {
-	int i;
-	int N=8,incX=0,incY=0;
+	blasint i;
+	blasint N=8,incX=0,incY=0;
 	double a=0.25;
 	double x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
 	double y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+
 	double x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
-	double y2[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	double y2[]={4.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
 
 	//OpenBLAS
 	BLASFUNC(daxpy)(&N,&a,x1,&incX,y1,&incY);
-	//reference
-	BLASFUNC_REF(daxpy)(&N,&a,x2,&incX,y2,&incY);
 
 	for(i=0; i<N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
 	}
 }
 
-void test_zaxpy_inc_0(void)
+CTEST(axpy,zaxpy_inc_0)
 {
-	int i;
-	int N=4,incX=0,incY=0;
+	blasint i;
+	blasint N=4,incX=0,incY=0;
 	double a[2]={0.25,0.5};
 	double x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
 	double y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
 	double x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
-	double y2[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	double y2[]={-3.0,9.0,6.0,8.0,2.0,4.0,6.0,8.0};
 
 	//OpenBLAS
 	BLASFUNC(zaxpy)(&N,a,x1,&incX,y1,&incY);
-	//reference
-	BLASFUNC_REF(zaxpy)(&N,a,x2,&incX,y2,&incY);
 
 	for(i=0; i<2*N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
 	}
 }
 
-void test_saxpy_inc_0(void)
+CTEST(axpy,saxpy_inc_0)
 {
-	int i;
-	int N=8,incX=0,incY=0;
+	blasint i;
+	blasint N=8,incX=0,incY=0;
 	float a=0.25;
 	float x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
 	float y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
 	float x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
-	float y2[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	float y2[]={4.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
 
 	//OpenBLAS
 	BLASFUNC(saxpy)(&N,&a,x1,&incX,y1,&incY);
-	//reference
-	BLASFUNC_REF(saxpy)(&N,&a,x2,&incX,y2,&incY);
 
 	for(i=0; i<N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
 	}
 }
 
-void test_caxpy_inc_0(void)
+CTEST(axpy,caxpy_inc_0)
 {
-	int i;
-	int N=4,incX=0,incY=0;
+	blasint i;
+	blasint N=4,incX=0,incY=0;
 	float a[2]={0.25,0.5};
 	float x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
 	float y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
 	float x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
-	float y2[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	float y2[]={-3.0,9.0,6.0,8.0,2.0,4.0,6.0,8.0};
 
 	//OpenBLAS
 	BLASFUNC(caxpy)(&N,a,x1,&incX,y1,&incY);
-	//reference
-	BLASFUNC_REF(caxpy)(&N,a,x2,&incX,y2,&incY);
 
 	for(i=0; i<2*N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
 	}
 }
diff --git a/utest/test_dotu.c b/utest/test_dotu.c
index 4ecc95915..37441a21a 100644
--- a/utest/test_dotu.c
+++ b/utest/test_dotu.c
@@ -31,46 +31,54 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 **********************************************************************************/
 
-#include "common_utest.h"
+#include "openblas_utest.h"
 #include <complex.h>
+#include <stdio.h>
 
-void test_zdotu_n_1(void)
+CTEST( zdotu,zdotu_n_1)
 {
-	int N=1,incX=1,incY=1;
+	blasint N=1,incX=1,incY=1;
 	double x1[]={1.0,1.0};
 	double y1[]={1.0,2.0};
-	double x2[]={1.0,1.0};
-	double y2[]={1.0,2.0};
-	double _Complex result1=0.0;
-	double _Complex result2=0.0;
-	//OpenBLAS
+	
+	openblas_complex_double result1=openblas_make_complex_double(0.0,0.0);
+        openblas_complex_double result2=openblas_make_complex_double(-1.0000,3.0000);
+#ifdef RETURN_BY_STACK
+	BLASFUNC(zdotu)(&result1,&N,x1,&incX,y1,&incY);
+#else
 	result1=BLASFUNC(zdotu)(&N,x1,&incX,y1,&incY);
-	//reference
-	result2=BLASFUNC_REF(zdotu)(&N,x2,&incX,y2,&incY);
-
-	CU_ASSERT_DOUBLE_EQUAL(creal(result1), creal(result2), CHECK_EPS);
-	CU_ASSERT_DOUBLE_EQUAL(cimag(result1), cimag(result2), CHECK_EPS);
-//	printf("\%lf,%lf\n",creal(result1),cimag(result1));
-
+#endif
+	
+#ifdef OPENBLAS_COMPLEX_STRUCT
+	ASSERT_DBL_NEAR_TOL(result1.real, result2.real, DOUBLE_EPS);
+	ASSERT_DBL_NEAR_TOL(result1.imag, result2.imag, DOUBLE_EPS);
+#else
+	ASSERT_DBL_NEAR_TOL(creal(result1), creal(result2), DOUBLE_EPS);
+	ASSERT_DBL_NEAR_TOL(cimag(result1), cimag(result2), DOUBLE_EPS);
+#endif
+	
 }
 
-void test_zdotu_offset_1(void)
+CTEST(zdotu, zdotu_offset_1)
 {
-	int N=1,incX=1,incY=1;
+	blasint N=1,incX=1,incY=1;
 	double x1[]={1.0,2.0,3.0,4.0};
 	double y1[]={5.0,6.0,7.0,8.0};
-	double x2[]={1.0,2.0,3.0,4.0};
-	double y2[]={5.0,6.0,7.0,8.0};
-	double _Complex result1=0.0;
-	double _Complex result2=0.0;
-	//OpenBLAS
+	
+	openblas_complex_double result1=openblas_make_complex_double(0.0,0.0);
+        openblas_complex_double result2=openblas_make_complex_double(-9.0,32.0);
+#ifdef RETURN_BY_STACK
+	BLASFUNC(zdotu)(&result1,&N,x1+1,&incX,y1+1,&incY);
+#else
 	result1=BLASFUNC(zdotu)(&N,x1+1,&incX,y1+1,&incY);
-	//reference
-	result2=BLASFUNC_REF(zdotu)(&N,x2+1,&incX,y2+1,&incY);
-
-	CU_ASSERT_DOUBLE_EQUAL(creal(result1), creal(result2), CHECK_EPS);
-	CU_ASSERT_DOUBLE_EQUAL(cimag(result1), cimag(result2), CHECK_EPS);
-//	printf("\%lf,%lf\n",creal(result1),cimag(result1));
+#endif
+	
+#ifdef OPENBLAS_COMPLEX_STRUCT
+	ASSERT_DBL_NEAR_TOL(result1.real, result2.real, DOUBLE_EPS);
+	ASSERT_DBL_NEAR_TOL(result1.imag, result2.imag, DOUBLE_EPS);
+#else
+	ASSERT_DBL_NEAR_TOL(creal(result1), creal(result2), DOUBLE_EPS);
+	ASSERT_DBL_NEAR_TOL(cimag(result1), cimag(result2), DOUBLE_EPS);
+#endif
 
 }
-
diff --git a/utest/test_dsdot.c b/utest/test_dsdot.c
index 536153c7e..7d082a372 100644
--- a/utest/test_dsdot.c
+++ b/utest/test_dsdot.c
@@ -31,21 +31,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 **********************************************************************************/
 
-#include "common_utest.h"
+#include "openblas_utest.h"
 
-void test_dsdot_n_1()
+CTEST(dsdot,dsdot_n_1)
 {
-	float x= 0.172555164;
-	float y= -0.0138700781;
-	int incx=1;
-	int incy=1;
-	int n=1;
+	float x= 0.172555164F;
+	float y= -0.0138700781F;
+	blasint incx=1;
+	blasint incy=1;
+	blasint n=1;
 
-	double res1=0.0f, res2=0.0f;
+	double res1=0.0f, res2=-0.00239335360107;
 
 	res1=BLASFUNC(dsdot)(&n, &x, &incx, &y, &incy);
-	res2=BLASFUNC_REF(dsdot)(&n, &x, &incx, &y, &incy);
-
-	CU_ASSERT_DOUBLE_EQUAL(res1, res2, CHECK_EPS);
+	ASSERT_DBL_NEAR_TOL(res1, res2, DOUBLE_EPS);
 
 }
diff --git a/utest/test_rot.c b/utest/test_rot.c
index b8f9f177a..2a47b8058 100644
--- a/utest/test_rot.c
+++ b/utest/test_rot.c
@@ -31,88 +31,82 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 **********************************************************************************/
 
-#include "common_utest.h"
+#include "openblas_utest.h"
 
-void test_drot_inc_0(void)
+CTEST(rot,drot_inc_0)
 {
-	int i=0;
-	int N=4,incX=0,incY=0;
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
 	double c=0.25,s=0.5;
 	double x1[]={1.0,3.0,5.0,7.0};
 	double y1[]={2.0,4.0,6.0,8.0};
-	double x2[]={1.0,3.0,5.0,7.0};
-	double y2[]={2.0,4.0,6.0,8.0};
+	double x2[]={-0.21484375000000,3.0,5.0,7.0};
+	double y2[]={ 0.03906250000000,4.0,6.0,8.0};
+
 
 	//OpenBLAS
 	BLASFUNC(drot)(&N,x1,&incX,y1,&incY,&c,&s);
-	//reference
-	BLASFUNC_REF(drot)(&N,x2,&incX,y2,&incY,&c,&s);
 
 	for(i=0; i<N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
 	}
 }
 
-void test_zdrot_inc_0(void)
+CTEST(rot,zdrot_inc_0)
 {
-	int i=0;
-	int N=4,incX=0,incY=0;
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
 	double c=0.25,s=0.5;
 	double x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
 	double y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
-	double x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
-	double y2[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	double x2[]={-0.21484375000000,-0.45703125000000 ,5.0,7.0,1.0,3.0,5.0,7.0};
+	double y2[]={ 0.03906250000000, 0.17187500000000 ,6.0,8.0,2.0,4.0,6.0,8.0};
+	
 
 	//OpenBLAS
 	BLASFUNC(zdrot)(&N,x1,&incX,y1,&incY,&c,&s);
-	//reference
-	BLASFUNC_REF(zdrot)(&N,x2,&incX,y2,&incY,&c,&s);
 
 	for(i=0; i<2*N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
 	}
 }
 
-void test_srot_inc_0(void)
+CTEST(rot,srot_inc_0)
 {
-	int i=0;
-	int N=4,incX=0,incY=0;
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
 	float c=0.25,s=0.5;
 	float x1[]={1.0,3.0,5.0,7.0};
 	float y1[]={2.0,4.0,6.0,8.0};
-	float x2[]={1.0,3.0,5.0,7.0};
-	float y2[]={2.0,4.0,6.0,8.0};
+	float x2[]={-0.21484375000000,3.0,5.0,7.0};
+	float y2[]={ 0.03906250000000,4.0,6.0,8.0};
 
 	//OpenBLAS
 	BLASFUNC(srot)(&N,x1,&incX,y1,&incY,&c,&s);
-	//reference
-	BLASFUNC_REF(srot)(&N,x2,&incX,y2,&incY,&c,&s);
 
 	for(i=0; i<N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], SINGLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], SINGLE_EPS);
 	}
 }
 
-void test_csrot_inc_0(void)
+CTEST(rot, csrot_inc_0)
 {
-	int i=0;
-	int N=4,incX=0,incY=0;
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
 	float c=0.25,s=0.5;
 	float x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
 	float y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
-	float x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
-	float y2[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
-
+	float x2[]={-0.21484375000000,-0.45703125000000 ,5.0,7.0,1.0,3.0,5.0,7.0};
+	float y2[]={ 0.03906250000000, 0.17187500000000 ,6.0,8.0,2.0,4.0,6.0,8.0};
+	
 	//OpenBLAS
 	BLASFUNC(csrot)(&N,x1,&incX,y1,&incY,&c,&s);
-	//reference
-	BLASFUNC_REF(csrot)(&N,x2,&incX,y2,&incY,&c,&s);
 
 	for(i=0; i<2*N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], SINGLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], SINGLE_EPS);
 	}
 }
diff --git a/utest/test_swap.c b/utest/test_swap.c
index cf70079fb..2f8b68ebd 100644
--- a/utest/test_swap.c
+++ b/utest/test_swap.c
@@ -31,32 +31,30 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 **********************************************************************************/
 
-#include "common_utest.h"
+#include "openblas_utest.h"
 
-void test_dswap_inc_0(void)
+CTEST(swap,dswap_inc_0)
 {
-	int i=0;
-	int N=4,incX=0,incY=0;
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
 	double x1[]={1.0,3.0,5.0,7.0};
 	double y1[]={2.0,4.0,6.0,8.0};
 	double x2[]={1.0,3.0,5.0,7.0};
 	double y2[]={2.0,4.0,6.0,8.0};
-
+	
 	//OpenBLAS
 	BLASFUNC(dswap)(&N,x1,&incX,y1,&incY);
-	//reference
-	BLASFUNC_REF(dswap)(&N,x2,&incX,y2,&incY);
 
 	for(i=0; i<N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
 	}
 }
 
-void test_zswap_inc_0(void)
+CTEST(swap,zswap_inc_0)
 {
-	int i=0;
-	int N=4,incX=0,incY=0;
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
 	double x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
 	double y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
 	double x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
@@ -64,19 +62,17 @@ void test_zswap_inc_0(void)
 
 	//OpenBLAS
 	BLASFUNC(zswap)(&N,x1,&incX,y1,&incY);
-	//reference
-	BLASFUNC_REF(zswap)(&N,x2,&incX,y2,&incY);
 
 	for(i=0; i<2*N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
 	}
 }
 
-void test_sswap_inc_0(void)
+CTEST(swap,sswap_inc_0)
 {
-	int i=0;
-	int N=4,incX=0,incY=0;
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
 	float x1[]={1.0,3.0,5.0,7.0};
 	float y1[]={2.0,4.0,6.0,8.0};
 	float x2[]={1.0,3.0,5.0,7.0};
@@ -84,19 +80,17 @@ void test_sswap_inc_0(void)
 
 	//OpenBLAS
 	BLASFUNC(sswap)(&N,x1,&incX,y1,&incY);
-	//reference
-	BLASFUNC_REF(sswap)(&N,x2,&incX,y2,&incY);
 
 	for(i=0; i<N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], SINGLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], SINGLE_EPS);
 	}
 }
 
-void test_cswap_inc_0(void)
+CTEST(swap,cswap_inc_0)
 {
-	int i=0;
-	int N=4,incX=0,incY=0;
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
 	float x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
 	float y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
 	float x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
@@ -104,11 +98,9 @@ void test_cswap_inc_0(void)
 
 	//OpenBLAS
 	BLASFUNC(cswap)(&N,x1,&incX,y1,&incY);
-	//reference
-	BLASFUNC_REF(cswap)(&N,x2,&incX,y2,&incY);
 
 	for(i=0; i<2*N; i++){
-		CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
-		CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], SINGLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], SINGLE_EPS);
 	}
 }
diff --git a/utest/utest_main2.c b/utest/utest_main2.c
index 51bfc40a2..bcaa43ec0 100644
--- a/utest/utest_main2.c
+++ b/utest/utest_main2.c
@@ -32,6 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 **********************************************************************************/
 
 #include <stdio.h>
+#include <complex.h>
 
 #define CTEST_MAIN
 #define CTEST_SEGFAULT
@@ -56,7 +57,7 @@ CTEST (drotmg,rotmg){
 	double te_y1, tr_y1;
 	double te_param[5];
 	double tr_param[5];
-	int i=0;
+	blasint i=0;
 	// original test case for libGoto bug fixed by feb2014 rewrite
 	te_d1= 0.21149573940783739;
 	te_d2= 0.046892057172954082;
@@ -103,7 +104,7 @@ CTEST (drotmg,rotmg_issue1452){
 	double te_y1, tr_y1;
 	double te_param[5];
 	double tr_param[5];
-	int i=0;
+	blasint i=0;
 
 	// from issue #1452, buggy version returned 0.000244 for param[3]
 	te_d1 = 5.9e-8;
@@ -148,7 +149,7 @@ CTEST(drotmg, rotmg_D1eqD2_X1eqX2){
 	double te_y1, tr_y1;
 	double te_param[5];
 	double tr_param[5];
-	int i=0;
+	blasint i=0;
 	te_d1= tr_d1=2.;
 	te_d2= tr_d2=2.;
 	te_x1= tr_x1=8.;
@@ -183,9 +184,314 @@ CTEST(drotmg, rotmg_D1eqD2_X1eqX2){
 	}
 }
 
+CTEST(axpy,daxpy_inc_0)
+{
+	blasint i;
+	blasint N=8,incX=0,incY=0;
+	double a=0.25;
+	double x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	double y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+
+	double x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	double y2[]={4.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+
+	//OpenBLAS
+	BLASFUNC(daxpy)(&N,&a,x1,&incX,y1,&incY);
+
+	for(i=0; i<N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
+	}
+}
+
+CTEST(axpy,zaxpy_inc_0)
+{
+	blasint i;
+	blasint N=4,incX=0,incY=0;
+	double a[2]={0.25,0.5};
+	double x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	double y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	double x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	double y2[]={-3.0,9.0,6.0,8.0,2.0,4.0,6.0,8.0};
+
+	//OpenBLAS
+	BLASFUNC(zaxpy)(&N,a,x1,&incX,y1,&incY);
+
+	for(i=0; i<2*N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
+	}
+}
+
+CTEST(axpy,saxpy_inc_0)
+{
+	blasint i;
+	blasint N=8,incX=0,incY=0;
+	float a=0.25;
+	float x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	float y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	float x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	float y2[]={4.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+
+	//OpenBLAS
+	BLASFUNC(saxpy)(&N,&a,x1,&incX,y1,&incY);
+
+	for(i=0; i<N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
+	}
+}
+
+CTEST(axpy,caxpy_inc_0)
+{
+	blasint i;
+	blasint N=4,incX=0,incY=0;
+	float a[2]={0.25,0.5};
+	float x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	float y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	float x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	float y2[]={-3.0,9.0,6.0,8.0,2.0,4.0,6.0,8.0};
+
+	//OpenBLAS
+	BLASFUNC(caxpy)(&N,a,x1,&incX,y1,&incY);
+
+	for(i=0; i<2*N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
+	}
+}
+
+CTEST( zdotu,zdotu_n_1)
+{
+	blasint N=1,incX=1,incY=1;
+	double x1[]={1.0,1.0};
+	double y1[]={1.0,2.0};
+        openblas_complex_double result1=openblas_make_complex_double(0.0,0.0);
+        openblas_complex_double result2=openblas_make_complex_double(-1.0,3.0);
+#ifdef RETURN_BY_STACK
+	BLASFUNC(zdotu)(&result1,&N,x1,&incX,y1,&incY);
+#else
+	result1=BLASFUNC(zdotu)(&N,x1,&incX,y1,&incY);
+#endif
+	
+#ifdef OPENBLAS_COMPLEX_STRUCT
+	ASSERT_DBL_NEAR_TOL(result1.real, result2.real, DOUBLE_EPS);
+	ASSERT_DBL_NEAR_TOL(result1.imag, result2.imag, DOUBLE_EPS);
+#else
+	ASSERT_DBL_NEAR_TOL(creal(result1), creal(result2), DOUBLE_EPS);
+	ASSERT_DBL_NEAR_TOL(cimag(result1), cimag(result2), DOUBLE_EPS);
+#endif
+}
+
+CTEST(zdotu, zdotu_offset_1)
+{
+	blasint N=1,incX=1,incY=1;
+	double x1[]={1.0,2.0,3.0,4.0};
+	double y1[]={5.0,6.0,7.0,8.0};
+        openblas_complex_double result1=openblas_make_complex_double(0.0,0.0);
+        openblas_complex_double result2=openblas_make_complex_double(-9.0,32.0);
+#ifdef RETURN_BY_STACK
+	BLASFUNC(zdotu)(&result1,&N,x1+1,&incX,y1+1,&incY);
+#else
+	result1=BLASFUNC(zdotu)(&N,x1+1,&incX,y1+1,&incY);
+#endif
+	
+#ifdef OPENBLAS_COMPLEX_STRUCT
+	ASSERT_DBL_NEAR_TOL(result1.real, result2.real, DOUBLE_EPS);
+	ASSERT_DBL_NEAR_TOL(result1.imag, result2.imag, DOUBLE_EPS);
+#else
+	ASSERT_DBL_NEAR_TOL(creal(result1), creal(result2), DOUBLE_EPS);
+	ASSERT_DBL_NEAR_TOL(cimag(result1), cimag(result2), DOUBLE_EPS);
+#endif
+}
+
+CTEST(dsdot,dsdot_n_1)
+{
+	float x= 0.172555164F;
+	float y= -0.0138700781F;
+	blasint incx=1;
+	blasint incy=1;
+	blasint n=1;
+
+	double res1=0.0f, res2=-0.00239335360107;
+
+	res1=BLASFUNC(dsdot)(&n, &x, &incx, &y, &incy);
+	ASSERT_DBL_NEAR_TOL(res1, res2, DOUBLE_EPS);
+
+}
+
+CTEST(rot,drot_inc_0)
+{
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
+	double c=0.25,s=0.5;
+	double x1[]={1.0,3.0,5.0,7.0};
+	double y1[]={2.0,4.0,6.0,8.0};
+	double x2[]={-0.21484375000000,3.0,5.0,7.0};
+	double y2[]={ 0.03906250000000,4.0,6.0,8.0};
+
+
+	//OpenBLAS
+	BLASFUNC(drot)(&N,x1,&incX,y1,&incY,&c,&s);
+
+	for(i=0; i<N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
+	}
+}
+
+CTEST(rot,zdrot_inc_0)
+{
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
+	double c=0.25,s=0.5;
+	double x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	double y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	double x2[]={-0.21484375000000,-0.45703125000000 ,5.0,7.0,1.0,3.0,5.0,7.0};
+	double y2[]={ 0.03906250000000, 0.17187500000000 ,6.0,8.0,2.0,4.0,6.0,8.0};
+	
+
+	//OpenBLAS
+	BLASFUNC(zdrot)(&N,x1,&incX,y1,&incY,&c,&s);
+
+	for(i=0; i<2*N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
+	}
+}
+
+CTEST(rot,srot_inc_0)
+{
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
+	float c=0.25,s=0.5;
+	float x1[]={1.0,3.0,5.0,7.0};
+	float y1[]={2.0,4.0,6.0,8.0};
+	float x2[]={-0.21484375000000,3.0,5.0,7.0};
+	float y2[]={ 0.03906250000000,4.0,6.0,8.0};
+
+	//OpenBLAS
+	BLASFUNC(srot)(&N,x1,&incX,y1,&incY,&c,&s);
+
+	for(i=0; i<N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], SINGLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], SINGLE_EPS);
+	}
+}
+
+CTEST(rot, csrot_inc_0)
+{
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
+	float c=0.25,s=0.5;
+	float x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	float y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	float x2[]={-0.21484375000000,-0.45703125000000 ,5.0,7.0,1.0,3.0,5.0,7.0};
+	float y2[]={ 0.03906250000000, 0.17187500000000 ,6.0,8.0,2.0,4.0,6.0,8.0};
+	
+	//OpenBLAS
+	BLASFUNC(csrot)(&N,x1,&incX,y1,&incY,&c,&s);
+
+	for(i=0; i<2*N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], SINGLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], SINGLE_EPS);
+	}
+}
+
+CTEST(swap,dswap_inc_0)
+{
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
+	double x1[]={1.0,3.0,5.0,7.0};
+	double y1[]={2.0,4.0,6.0,8.0};
+	double x2[]={1.0,3.0,5.0,7.0};
+	double y2[]={2.0,4.0,6.0,8.0};
+	
+	//OpenBLAS
+	BLASFUNC(dswap)(&N,x1,&incX,y1,&incY);
+
+	for(i=0; i<N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
+	}
+}
+
+CTEST(swap,zswap_inc_0)
+{
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
+	double x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	double y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	double x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	double y2[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+
+	//OpenBLAS
+	BLASFUNC(zswap)(&N,x1,&incX,y1,&incY);
+
+	for(i=0; i<2*N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], DOUBLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], DOUBLE_EPS);
+	}
+}
+
+CTEST(swap,sswap_inc_0)
+{
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
+	float x1[]={1.0,3.0,5.0,7.0};
+	float y1[]={2.0,4.0,6.0,8.0};
+	float x2[]={1.0,3.0,5.0,7.0};
+	float y2[]={2.0,4.0,6.0,8.0};
+
+	//OpenBLAS
+	BLASFUNC(sswap)(&N,x1,&incX,y1,&incY);
+
+	for(i=0; i<N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], SINGLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], SINGLE_EPS);
+	}
+}
+
+CTEST(swap,cswap_inc_0)
+{
+	blasint i=0;
+	blasint N=4,incX=0,incY=0;
+	float x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	float y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+	float x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
+	float y2[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
+
+	//OpenBLAS
+	BLASFUNC(cswap)(&N,x1,&incX,y1,&incY);
+
+	for(i=0; i<2*N; i++){
+		ASSERT_DBL_NEAR_TOL(x1[i], x2[i], SINGLE_EPS);
+		ASSERT_DBL_NEAR_TOL(y1[i], y2[i], SINGLE_EPS);
+	}
+}
+
 int main(int argc, const char ** argv){
 
   CTEST_ADD(amax, samax);
+  CTEST_ADD (drotmg,rotmg);
+  CTEST_ADD (drotmg,rotmg_issue1452);
+  CTEST_ADD (drotmg, rotmg_D1eqD2_X1eqX2);
+  CTEST_ADD (axpy,daxpy_inc_0);
+  CTEST_ADD (axpy,zaxpy_inc_0);
+  CTEST_ADD (axpy,saxpy_inc_0);
+  CTEST_ADD (axpy,caxpy_inc_0);
+  CTEST_ADD (zdotu,zdotu_n_1);
+  CTEST_ADD (zdotu, zdotu_offset_1);
+  CTEST_ADD (dsdot,dsdot_n_1);
+  CTEST_ADD (rot,drot_inc_0);
+  CTEST_ADD (rot,zdrot_inc_0);
+  CTEST_ADD (rot,srot_inc_0);
+  CTEST_ADD (rot, csrot_inc_0);
+  CTEST_ADD (swap,dswap_inc_0);
+  CTEST_ADD (swap,zswap_inc_0);
+  CTEST_ADD (swap,sswap_inc_0);
+  CTEST_ADD (swap,cswap_inc_0);
+
   int num_fail=0;
 
   num_fail=ctest_main(argc, argv);