diff --git a/paddle/cuda/include/hl_tensor_ops.h b/paddle/cuda/include/hl_tensor_ops.h
index 93d38b7d2299d994cde0934213668a525bffa80c..b2bf334dab9799153fe1d4fe2c74cce9d57168b9 100644
--- a/paddle/cuda/include/hl_tensor_ops.h
+++ b/paddle/cuda/include/hl_tensor_ops.h
@@ -461,7 +461,7 @@ class add<float32x4_t> {
 public:
   INLINE float32x4_t operator()(const float32x4_t a,
                                 const float32x4_t b) const {
-    return vmulq_f32(a, b);
+    return vaddq_f32(a, b);
   }
 };