diff --git a/mace/ops/arm/fp16/gemv.h b/mace/ops/arm/fp16/gemv.h
index 8e7e2a3c91438303f6724b740f16739db2ed5ebc..aa5add8cf684a13c1a036c8784eda10298b163f0 100644
--- a/mace/ops/arm/fp16/gemv.h
+++ b/mace/ops/arm/fp16/gemv.h
@@ -15,6 +15,13 @@
 #ifndef MACE_OPS_ARM_FP16_GEMV_H_
 #define MACE_OPS_ARM_FP16_GEMV_H_
 
+#if defined(MACE_ENABLE_NEON) && \
+    defined(__ARM_FP16_FORMAT_IEEE) && (__ARM_FP & 2)
+// TODO(lichao): replace it with global macro
+#define MACE_ENABLE_FP16_NEON
+#endif
+
+
 #include "mace/core/types.h"
 
 #if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
@@ -37,7 +44,7 @@ void FP16Gemv(const INPUT_TYPE_LEFT *m_ptr,
               const index_t width,
               OUTPUT_TYPE *result);
 
-#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
+#if defined(MACE_ENABLE_FP16_NEON) && defined(__ANDROID__)
 template<>
 void FP16Gemv<float16_t, float, float>(const float16_t *m_ptr,
                                        const float *v_ptr,
@@ -112,7 +119,7 @@ void FP16Gemv<float16_t, float, float>(const float16_t *m_ptr,
     *out_ptr0++ = sum0;
   }
 }
-#endif
+#endif  // MACE_ENABLE_FP16_NEON && __ANDROID__
 
 }  // namespace ops
 }  // namespace mace
diff --git a/mace/ops/matmul.cc b/mace/ops/matmul.cc
index 394ace009931cb000ede82606b5a60b46edce35d..1c97279e90f3ccd5792c1ea866729ef0842b9bb4 100644
--- a/mace/ops/matmul.cc
+++ b/mace/ops/matmul.cc
@@ -513,7 +513,7 @@ class MatMulOp<DeviceType::GPU, float> : public MatMulOpBase {
 };
 #endif  // MACE_ENABLE_OPENCL
 
-#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
+#if defined(MACE_ENABLE_FP16_NEON) && defined(__ANDROID__)
 template <>
 class MatMulOp<CPU, float16_t> : public MatMulOpBase {
  public:
@@ -590,7 +590,7 @@ class MatMulOp<CPU, float16_t> : public MatMulOpBase {
 
  private:
 };
-#endif  // MACE_ENABLE_NEON
+#endif  // MACE_ENABLE_FP16_NEON
 
 void RegisterMatMul(OpRegistryBase *op_registry) {
   MACE_REGISTER_OP(op_registry, "MatMul", MatMulOp,
@@ -601,10 +601,10 @@ void RegisterMatMul(OpRegistryBase *op_registry) {
                    DeviceType::CPU, uint8_t);
 #endif  // MACE_ENABLE_QUANTIZE
 
-#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
+#if defined(MACE_ENABLE_FP16_NEON) && defined(__ANDROID__)
   MACE_REGISTER_OP(op_registry, "MatMul", MatMulOp,
                    DeviceType::CPU, float16_t);
-#endif  // MACE_ENABLE_NEON
+#endif  // MACE_ENABLE_FP16_NEON && __ANDROID__
 }
 
 }  // namespace ops
diff --git a/tools/device.py b/tools/device.py
index dc51e4aec22e02e034e8849302bdeffe48a51bda..55cf3a75f77cc0343698d0cf1128cbef0044774f 100644
--- a/tools/device.py
+++ b/tools/device.py
@@ -271,7 +271,7 @@ class DeviceWrapper:
 
             if device_type == common.DeviceType.GPU:
                 if os.path.exists(opencl_binary_file):
-                    self.push(opencl_binary_file, self.data_dir)
+                    self.push(opencl_binary_file, self.interior_dir)
                 if os.path.exists(opencl_parameter_file):
                     self.push(opencl_parameter_file, self.data_dir)