diff --git a/tests/opencl/Makefile b/tests/opencl/Makefile
index c838c3de..acb96ba9 100644
--- a/tests/opencl/Makefile
+++ b/tests/opencl/Makefile
@@ -36,9 +36,9 @@ run-simx:
 	$(MAKE) -C lbm run-simx
 	$(MAKE) -C oclprintf run-simx
 	$(MAKE) -C blackscholes run-simx
-	$(MAKE) -C matmul run-simx
 	$(MAKE) -C transpose run-simx
 	$(MAKE) -C convolution run-simx
+#	$(MAKE) -C matmul run-simx
 #	$(MAKE) -C vectorhypot run-simx	
 #	$(MAKE) -C mri-q run-simx
 
@@ -59,8 +59,8 @@ run-rtlsim:
 	$(MAKE) -C lbm run-rtlsim
 	$(MAKE) -C oclprintf run-rtlsim
 	$(MAKE) -C blackscholes run-rtlsim
-	$(MAKE) -C matmul run-rtlsim
 	$(MAKE) -C convolution run-rtlsim
+#	$(MAKE) -C matmul run-rtlsim
 #	$(MAKE) -C vectorhypot run-rtlsim
 #	$(MAKE) -C mri-q run-rtlsim
 
@@ -81,8 +81,8 @@ run-opae:
 	$(MAKE) -C lbm run-opae
 	$(MAKE) -C oclprintf run-opae
 	$(MAKE) -C blackscholes run-opae
-	$(MAKE) -C matmul run-opae
 	$(MAKE) -C convolution run-opae
+#	$(MAKE) -C matmul run-opae
 #	$(MAKE) -C vectorhypot run-opae
 #	$(MAKE) -C mri-q run-opae
 
diff --git a/tests/opencl/common.mk b/tests/opencl/common.mk
index ce82dac3..762712b4 100644
--- a/tests/opencl/common.mk
+++ b/tests/opencl/common.mk
@@ -41,13 +41,12 @@ CXXFLAGS += -std=c++11 -Wall -Wextra -Wfatal-errors
 CXXFLAGS += -Wno-deprecated-declarations -Wno-unused-parameter -Wno-narrowing
 CXXFLAGS += -pthread
 CXXFLAGS += -I$(POCL_RT_PATH)/include
-LDFLAGS  += -L$(POCL_RT_PATH)/lib -L$(VORTEX_RT_PATH)/stub -lvortex
 
 ifdef HOSTGPU
 	CXXFLAGS += -DHOSTGPU
 	LDFLAGS += -lOpenCL
 else
-	LDFLAGS += $(POCL_RT_PATH)/lib/libOpenCL.so	
+	LDFLAGS += -L$(VORTEX_RT_PATH)/stub -lvortex $(POCL_RT_PATH)/lib/libOpenCL.so
 endif
 
 # Debugigng
diff --git a/tests/opencl/oclprintf/main.cc b/tests/opencl/oclprintf/main.cc
index 184eec96..4af39802 100644
--- a/tests/opencl/oclprintf/main.cc
+++ b/tests/opencl/oclprintf/main.cc
@@ -106,11 +106,6 @@ int main (int argc, char **argv) {
   
   cl_platform_id platform_id;
   size_t kernel_size;
-  cl_int binary_status;
-
-  // read kernel binary from file  
-  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
-    return -1;
   
   // Getting platform and device information
   CL_CHECK(clGetPlatformIDs(1, &platform_id, NULL));
@@ -124,12 +119,17 @@ int main (int argc, char **argv) {
   a_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_READ_ONLY, nbytes, NULL, &_err));
 
   printf("Create program from kernel source\n");
-  program = CL_CHECK2(clCreateProgramWithBinary(
-    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, &binary_status, &_err));
-  if (program == NULL) {
-    cleanup();
+#ifdef HOSTGPU
+  if (0 != read_kernel_file("kernel.cl", &kernel_bin, &kernel_size))
     return -1;
-  }
+  program = CL_CHECK2(clCreateProgramWithSource(
+    context, 1, (const char**)&kernel_bin, &kernel_size, &_err));  
+#else
+  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
+    return -1;
+  program = CL_CHECK2(clCreateProgramWithBinary(
+    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, NULL, &_err));
+#endif
 
   // Build program
   CL_CHECK(clBuildProgram(program, 1, &device_id, NULL, NULL, NULL));
diff --git a/tests/opencl/psort/main.cc b/tests/opencl/psort/main.cc
index b627ceee..8bc834dc 100644
--- a/tests/opencl/psort/main.cc
+++ b/tests/opencl/psort/main.cc
@@ -115,11 +115,6 @@ int main (int argc, char **argv) {
   
   cl_platform_id platform_id;
   size_t kernel_size;
-  cl_int binary_status;
-
-  // read kernel binary from file  
-  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
-    return -1;
   
   // Getting platform and device information
   CL_CHECK(clGetPlatformIDs(1, &platform_id, NULL));
@@ -134,12 +129,17 @@ int main (int argc, char **argv) {
   c_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_WRITE_ONLY, nbytes, NULL, &_err));
 
   printf("Create program from kernel source\n");
-  program = CL_CHECK2(clCreateProgramWithBinary(
-    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, &binary_status, &_err));
-  if (program == NULL) {
-    cleanup();
+#ifdef HOSTGPU
+  if (0 != read_kernel_file("kernel.cl", &kernel_bin, &kernel_size))
     return -1;
-  }
+  program = CL_CHECK2(clCreateProgramWithSource(
+    context, 1, (const char**)&kernel_bin, &kernel_size, &_err));  
+#else
+  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
+    return -1;
+  program = CL_CHECK2(clCreateProgramWithBinary(
+    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, NULL, &_err));
+#endif
 
   // Build program
   CL_CHECK(clBuildProgram(program, 1, &device_id, NULL, NULL, NULL));
@@ -159,12 +159,12 @@ int main (int argc, char **argv) {
   for (int i = 0; i < size; ++i) {
     if (float_enable) {
       float value = sinf(i)*sinf(i);
-      h_a[i] = *(int*)&value;
-      printf("*** [%d]: h_a=%f\n", i, value);
+      ((float*)h_a)[i] = value;
+      printf("*** [%d]: %f\n", i, value);
     } else {      
       int value = size*sinf(i);
       h_a[i] = value;
-      printf("*** [%d]: h_a=%d\n", i, value);
+      printf("*** [%d]: %d\n", i, value);
     }
   }
 
@@ -188,38 +188,44 @@ int main (int argc, char **argv) {
   CL_CHECK(clEnqueueReadBuffer(commandQueue, c_memobj, CL_TRUE, 0, nbytes, h_c, 0, NULL, NULL));
 
   printf("Verify result\n");
-  for (int i = 0; i < size; ++i) {
-    int value = h_c[i];
+  for (int i = 0; i < size; ++i) {    
     if (float_enable) {
-      printf("*** [%d]: h_a=%f\n", i, *(float*)&value);
+      float value = ((float*)h_c)[i];
+      printf("*** [%d]: %f\n", i, value);
     } else {
-      printf("*** [%d]: h_a=%d\n", i, value);
+      int value = h_c[i];
+      printf("*** [%d]: %d\n", i, value);
     }
   }
   int errors = 0;
-  for (int i = 0; i < size; ++i) {
-    int ref = h_a[i];
-    float ref_f = *(float*)&ref;
+  for (int i = 0; i < size; ++i) {    
     int pos = 0;
-    for (int j = 0; j < size; ++j) {
-      int cur = h_a[j];
-      if (float_enable) {
-        float cur_f = *(float*)&cur;
-        pos += (cur_f < ref_f) || (cur_f == ref_f && j < i);
-      } else {
+    if (float_enable) {    
+      float ref = ((float*)h_a)[i];
+      for (int j = 0; j < size; ++j) {        
+        float cur = ((float*)h_a)[j];
+        pos += (cur < ref) || (cur == ref && j < i);
+      }        
+      float value = ((float*)h_c)[pos];
+      if (value != ref) {
+        if (errors < 100) {        
+          printf("*** error: [%d] expected=%f, actual=%f\n", pos, ref, value);
+        }        
+        ++errors;
+      }
+    } else {      
+      int ref = h_a[i];
+      for (int j = 0; j < size; ++j) {        
+        int cur = h_a[j];
         pos += (cur < ref) || (cur == ref && j < i);
       }
-    }
-    int value = h_c[pos];
-    if (value != ref) {
-      if (errors < 100) {
-        if (float_enable) {
-          printf("*** error: [%d] expected=%f, actual=%f\n", pos, ref_f, *(float*)&value);
-        } else {
+      int value = h_c[pos];
+      if (value != ref) {
+        if (errors < 100) {        
           printf("*** error: [%d] expected=%d, actual=%d\n", pos, ref, value);
         }
+        ++errors;
       }
-      ++errors;
     }
   }
   if (0 == errors) {
diff --git a/tests/opencl/saxpy/main.cc b/tests/opencl/saxpy/main.cc
index 4ea15759..9355c945 100644
--- a/tests/opencl/saxpy/main.cc
+++ b/tests/opencl/saxpy/main.cc
@@ -151,16 +151,12 @@ int main(int argc, char **argv) {
   
   cl_platform_id platform_id;
   cl_device_id device_id;
+  cl_program program;
   cl_mem input_buffer;
   cl_mem output_buffer;
   size_t kernel_size;
   cl_context context;
   cl_command_queue queue;
-  cl_int binary_status = 0;
-
-  // read kernel binary from file  
-  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
-    return -1;
 
   // Getting platform and device information
   CL_CHECK(clGetPlatformIDs(1, &platform_id, NULL));
@@ -172,19 +168,18 @@ int main(int argc, char **argv) {
   cl_kernel kernel = 0;
   cl_mem memObjects[2] = {0, 0};
 
-  // Create OpenCL program - first attempt to load cached binary.
-  //  If that is not available, then create the program from source
-  //  and store the binary for future use.
-  printf("create program from binary...\n");
-  cl_program program = CL_CHECK_ERR(clCreateProgramWithBinary(
-    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, &binary_status, &_err));
-  if (program == NULL) {
-    std::cerr << "Failed to write program binary" << std::endl;
-    Cleanup(device_id, context, queue, program, kernel, memObjects);
-    return 1;
-  } else {
-    printf("Read program from binary.\n");
-  }
+  printf("Create program from kernel source\n");
+#ifdef HOSTGPU
+  if (0 != read_kernel_file("kernel.cl", &kernel_bin, &kernel_size))
+    return -1;
+  program = CL_CHECK_ERR(clCreateProgramWithSource(
+    context, 1, (const char**)&kernel_bin, &kernel_size, &_err));  
+#else
+  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
+    return -1;
+  program = CL_CHECK_ERR(clCreateProgramWithBinary(
+    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, NULL, &_err));
+#endif
 
   // Build program
   CL_CHECK(clBuildProgram(program, 1, &device_id, NULL, NULL, NULL));
diff --git a/tests/opencl/sfilter/main.cc b/tests/opencl/sfilter/main.cc
index 3a7a5979..0ae264a5 100644
--- a/tests/opencl/sfilter/main.cc
+++ b/tests/opencl/sfilter/main.cc
@@ -149,14 +149,10 @@ int main(int argc, char **argv) {
 
   cl_platform_id platform_id;
   cl_device_id device_id;
+  cl_program program;
   size_t kernel_size;
-  cl_int binary_status = 0;
   uint8_t *kernel_bin = NULL;
 
-  // read kernel binary from file  
-  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
-    return -1;
-
   // Getting platform and device information
   CL_CHECK(clGetPlatformIDs(1, &platform_id, NULL));
   CL_CHECK(clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_DEFAULT, 1, &device_id, NULL));
@@ -170,19 +166,18 @@ int main(int argc, char **argv) {
   cl_kernel kernel = 0;
   cl_mem memObjects[2] = {0, 0};
 
-  // Create OpenCL program - first attempt to load cached binary.
-  // If that is not available, then create the program from source
-  // and store the binary for future use.
-  printf("create program from binary...\n");
-  cl_program program = CL_CHECK_ERR(clCreateProgramWithBinary(
-    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, &binary_status, &_err));
-  if (program == NULL) {
-    std::cerr << "Failed to write program binary" << std::endl;
-    Cleanup(kernel_bin, device_id, context, queue, program, kernel, memObjects);
-    return 1;
-  } else {
-    printf("Read program from binary.");
-  }
+  printf("Create program from kernel source\n");
+#ifdef HOSTGPU
+  if (0 != read_kernel_file("kernel.cl", &kernel_bin, &kernel_size))
+    return -1;
+  program = CL_CHECK_ERR(clCreateProgramWithSource(
+    context, 1, (const char**)&kernel_bin, &kernel_size, &_err));  
+#else
+  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
+    return -1;
+  program = CL_CHECK_ERR(clCreateProgramWithBinary(
+    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, NULL, &_err));
+#endif
 
   // Build program
   CL_CHECK(clBuildProgram(program, 1, &device_id, NULL, NULL, NULL));
diff --git a/tests/opencl/sgemm/main.cc b/tests/opencl/sgemm/main.cc
index 7a02929f..bc48dff0 100644
--- a/tests/opencl/sgemm/main.cc
+++ b/tests/opencl/sgemm/main.cc
@@ -185,13 +185,8 @@ int main (int argc, char **argv) {
 
   cl_platform_id platform_id;
   size_t kernel_size;
-  cl_int binary_status;
 
   srand(50);
-
-  // read kernel binary from file  
-  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
-    return -1;
   
   // Getting platform and device information
   CL_CHECK(clGetPlatformIDs(1, &platform_id, NULL));
@@ -207,12 +202,17 @@ int main (int argc, char **argv) {
   c_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_WRITE_ONLY, nbytes, NULL, &_err));
 
   printf("Create program from kernel source\n");
-  program = CL_CHECK2(clCreateProgramWithBinary(
-    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, &binary_status, &_err));
-  if (program == NULL) {
-    cleanup();
+#ifdef HOSTGPU
+  if (0 != read_kernel_file("kernel.cl", &kernel_bin, &kernel_size))
     return -1;
-  }
+  program = CL_CHECK2(clCreateProgramWithSource(
+    context, 1, (const char**)&kernel_bin, &kernel_size, &_err));  
+#else
+  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
+    return -1;
+  program = CL_CHECK2(clCreateProgramWithBinary(
+    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, NULL, &_err));
+#endif
 
   // Build program
   CL_CHECK(clBuildProgram(program, 1, &device_id, NULL, NULL, NULL));
diff --git a/tests/opencl/vecadd/main.cc b/tests/opencl/vecadd/main.cc
index 992e88be..e443f7c5 100644
--- a/tests/opencl/vecadd/main.cc
+++ b/tests/opencl/vecadd/main.cc
@@ -122,11 +122,6 @@ int main (int argc, char **argv) {
   
   cl_platform_id platform_id;
   size_t kernel_size;
-  cl_int binary_status;
-
-  // read kernel binary from file  
-  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
-    return -1;
   
   // Getting platform and device information
   CL_CHECK(clGetPlatformIDs(1, &platform_id, NULL));
@@ -142,13 +137,17 @@ int main (int argc, char **argv) {
   c_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_WRITE_ONLY, nbytes, NULL, &_err));
 
   printf("Create program from kernel source\n");
-  cl_int _err;
-  program = clCreateProgramWithBinary(
-    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, &binary_status, &_err);
-  if (program == NULL) {
-    cleanup();
+#ifdef HOSTGPU
+  if (0 != read_kernel_file("kernel.cl", &kernel_bin, &kernel_size))
     return -1;
-  }
+  program = CL_CHECK2(clCreateProgramWithSource(
+    context, 1, (const char**)&kernel_bin, &kernel_size, &_err));  
+#else
+  if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
+    return -1;
+  program = CL_CHECK2(clCreateProgramWithBinary(
+    context, 1, &device_id, &kernel_size, (const uint8_t**)&kernel_bin, NULL, &_err));
+#endif
 
   // Build program
   CL_CHECK(clBuildProgram(program, 1, &device_id, NULL, NULL, NULL));