diff --git a/paddle/math/float16.h b/paddle/math/float16.h
index 3275546e69ceb503fb6ae61725a971cc3e8341d8..6799a83bd391ab722d8bf1a08cdd9118133bfdcb 100644
--- a/paddle/math/float16.h
+++ b/paddle/math/float16.h
@@ -118,8 +118,8 @@ struct PADDLE_ALIGN(2) float16 {
   PADDLE_HOSTDEVICE inline float16(const Eigen::half& h) : x(h.x) {}
 #endif  // USE_EIGEN
 
-#if (PADDLE_GNUC_VER >= 61 || PADDLE_CLANG_VER >= 34) && \
-    defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16)
+#if defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16) && \
+    (PADDLE_GNUC_VER >= 61 || PADDLE_CLANG_VER >= 34)
   // __fp16 is a native half precision data type for arm cpu,
   // float16_t is an alias for __fp16 in arm_fp16.h,
   // which is included in arm_neon.h.
@@ -207,8 +207,8 @@ struct PADDLE_ALIGN(2) float16 {
   }
 #endif  // USE_EIGEN
 
-#if (PADDLE_GNUC_VER >= 61 || PADDLE_CLANG_VER >= 34) && \
-    defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16)
+#if defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16) && \
+    (PADDLE_GNUC_VER >= 61 || PADDLE_CLANG_VER >= 34)
   PADDLE_HOSTDEVICE inline float16& operator=(const float16_t* rhs) {
     x = *reinterpret_cast<uint16_t*>(rhs);
     return *this;
@@ -302,8 +302,8 @@ struct PADDLE_ALIGN(2) float16 {
   }
 #endif  // USE_EIGEN
 
-#if (PADDLE_GNUC_VER >= 61 || PADDLE_CLANG_VER >= 34) && \
-    defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16)
+#if defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16) && \
+    (PADDLE_GNUC_VER >= 61 || PADDLE_CLANG_VER >= 34)
   // check whether it works or not
   PADDLE_HOSTDEVICE inline operator float16_t() const {
     float16 h = *this;
diff --git a/paddle/math/tests/CMakeLists.txt b/paddle/math/tests/CMakeLists.txt
index ab4ac38b3c5aea5239b2b52efe20dffefa62dba1..dc06f990905459d5b63d21596b3e8d2aae603fef 100644
--- a/paddle/math/tests/CMakeLists.txt
+++ b/paddle/math/tests/CMakeLists.txt
@@ -22,15 +22,18 @@ if(WITH_GPU)
     link_paddle_test(test_Tensor)
     CUDA_ADD_EXECUTABLE(test_lazyAssign test_lazyAssign.cu)
     link_paddle_test(test_lazyAssign) 
+    CUDA_ADD_EXECUTABLE(test_float16 test_float16.cu)
+    link_paddle_test(test_float16)
 else()
     compile_cu_as_cpp(test_Tensor.cu)
     add_unittest(test_Tensor test_Tensor.cu)
     compile_cu_as_cpp(test_lazyAssign.cu)
     add_unittest(test_lazyAssign test_lazyAssign.cu)
+    compile_cu_as_cpp(test_float16.cu)
+    add_unittest(test_float16 test_float16.cu)
 endif(WITH_GPU)
 
 add_simple_unittest(test_FPException)
 add_simple_unittest(test_GpuProfiler)
 add_simple_unittest(test_BaseMatrix)
 add_simple_unittest(test_Matrix)
-add_simple_unittest(test_float16)
diff --git a/paddle/math/tests/test_float16.cpp b/paddle/math/tests/test_float16.cu
similarity index 98%
rename from paddle/math/tests/test_float16.cpp
rename to paddle/math/tests/test_float16.cu
index 79f63d3a806ffb016d89368d665dc6bcea885080..40bc54f5b46e0111d10566e5978d338d90e24150 100644
--- a/paddle/math/tests/test_float16.cpp
+++ b/paddle/math/tests/test_float16.cu
@@ -15,7 +15,7 @@ limitations under the License. */
 namespace paddle {
 
 #ifdef PADDLE_CUDA_FP16
-TEST(float16, gpu) {
+TEST(float16, conversion_gpu) {
   // Conversion to and from cuda half
   float16 v1 = half(float16(1.0f));
   EXPECT_EQ(v1.x, 0x3c00);