diff --git a/driver/common/vx_utils.cpp b/driver/common/vx_utils.cpp
index 885da189..79853aa1 100644
--- a/driver/common/vx_utils.cpp
+++ b/driver/common/vx_utils.cpp
@@ -34,7 +34,7 @@ extern int vx_upload_kernel_bytes(vx_device_h device, const void* content, size_
     auto chunk_size = std::min<size_t>(buffer_transfer_size, size - offset);
     std::memcpy(buf_ptr, (uint8_t*)content + offset, chunk_size);
 
-    /*printf("**  Upload Kernel to 0x%0x: data=", kernel_base_addr + offset);
+    /*printf("***  Upload Kernel to 0x%0x: data=", kernel_base_addr + offset);
     for (int i = 0, n = ((chunk_size+7)/8); i < n; ++i) {
       printf("%08x", ((uint64_t*)((uint8_t*)content + offset))[n-1-i]);
     }
diff --git a/tests/opencl/Makefile b/tests/opencl/Makefile
index e5bde6c1..91aad370 100644
--- a/tests/opencl/Makefile
+++ b/tests/opencl/Makefile
@@ -6,6 +6,7 @@ all:
 	$(MAKE) -C nearn
 	$(MAKE) -C guassian
 	$(MAKE) -C printf
+	$(MAKE) -C psort
 	
 run:
 	$(MAKE) -C vecadd run-vlsim 
@@ -15,6 +16,7 @@ run:
 	$(MAKE) -C nearn run-vlsim
 	$(MAKE) -C guassian run-vlsim
 	$(MAKE) -C printf run-vlsim
+#$(MAKE) -C psort run-vlsim
 
 clean:
 	$(MAKE) -C vecadd clean
@@ -24,6 +26,7 @@ clean:
 	$(MAKE) -C nearn clean
 	$(MAKE) -C guassian clean
 	$(MAKE) -C printf clean
+	$(MAKE) -C psort clean
 
 clean-all:
 	$(MAKE) -C vecadd clean-all
@@ -32,4 +35,5 @@ clean-all:
 	$(MAKE) -C sfilter clean-all
 	$(MAKE) -C nearn clean-all
 	$(MAKE) -C guassian clean-all
-	$(MAKE) -C printf clean-all
\ No newline at end of file
+	$(MAKE) -C printf clean-all
+	$(MAKE) -C psort clean-all
\ No newline at end of file
diff --git a/tests/regression/io_addr/main.cpp b/tests/regression/io_addr/main.cpp
index 4c43838c..7899aa2a 100644
--- a/tests/regression/io_addr/main.cpp
+++ b/tests/regression/io_addr/main.cpp
@@ -183,9 +183,10 @@ int main(int argc, char *argv[]) {
   
   // allocate shared memory  
   std::cout << "allocate shared memory" << std::endl;    
-  uint32_t staging_buf_size = std::max<uint32_t>(src_buf_size,
+  uint32_t staging_buf_size = std::max<uint32_t>(NUM_ADDRS * sizeof(uint32_t),
+                                std::max<uint32_t>(src_buf_size,
                                   std::max<uint32_t>(dst_buf_size, 
-                                    sizeof(kernel_arg_t)));
+                                    sizeof(kernel_arg_t))));
   RT_CHECK(vx_alloc_shared_mem(device, staging_buf_size, &staging_buf));
   
   // upload kernel argument