diff --git a/paddle/fluid/inference/utils/benchmark.cc b/paddle/fluid/inference/utils/benchmark.cc
index 021edc2de5e90023fcd1431dd2025450e7462bd9..d03aa11b75ee58524746212e43a5796773f47932 100644
--- a/paddle/fluid/inference/utils/benchmark.cc
+++ b/paddle/fluid/inference/utils/benchmark.cc
@@ -33,7 +33,7 @@ std::string Benchmark::SerializeToString() const {
   ss << batch_size_ << "\t";
   ss << num_threads_ << "\t";
   ss << latency_ << "\t";
-  ss << 1000 / latency_;
+  ss << 1000.0 / latency_;
   ss << '\n';
   return ss.str();
 }
diff --git a/paddle/fluid/inference/utils/benchmark.h b/paddle/fluid/inference/utils/benchmark.h
index 80e8f77adb4ff2cc81a2a3dd0c44e4e304800122..76a3dd2c2992ebdf2528c539b3d161f558b34a08 100644
--- a/paddle/fluid/inference/utils/benchmark.h
+++ b/paddle/fluid/inference/utils/benchmark.h
@@ -11,9 +11,11 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
+#pragma once
 
 #include <fstream>
 #include <iostream>
+#include <string>
 
 namespace paddle {
 namespace inference {
@@ -31,8 +33,8 @@ struct Benchmark {
   bool use_gpu() const { return use_gpu_; }
   void SetUseGpu() { use_gpu_ = true; }
 
-  int latency() const { return latency_; }
-  void SetLatency(int x) { latency_ = x; }
+  float latency() const { return latency_; }
+  void SetLatency(float x) { latency_ = x; }
 
   const std::string& name() const { return name_; }
   void SetName(const std::string& name) { name_ = name; }
@@ -43,7 +45,7 @@ struct Benchmark {
  private:
   bool use_gpu_{false};
   int batch_size_{0};
-  int latency_;
+  float latency_;
   int num_threads_{1};
   std::string name_;
 };