diff --git a/benchmark/.gitignore b/benchmark/.gitignore
deleted file mode 100644
index 89096b6cd..000000000
--- a/benchmark/.gitignore
+++ /dev/null
@@ -1,13 +0,0 @@
-*.tgz
-*.zip
-*.tar
-*.tar.gz
-*.tgz
-*.jpg
-*.png
-*.jpeg
-*.txt
-*.log
-yolov8_s_*
-._yolov8_s_*
-Mobile*
\ No newline at end of file
diff --git a/benchmark/cpp/CMakeLists.txt b/benchmark/cpp/CMakeLists.txt
new file mode 100755
index 000000000..9706587d3
--- /dev/null
+++ b/benchmark/cpp/CMakeLists.txt
@@ -0,0 +1,17 @@
+PROJECT(infer_demo C CXX)
+CMAKE_MINIMUM_REQUIRED (VERSION 3.10)
+
+# specify the decompress directory of FastDeploy SDK
+option(FASTDEPLOY_INSTALL_DIR "Path of downloaded fastdeploy sdk.")
+include(${FASTDEPLOY_INSTALL_DIR}/utils/gflags.cmake)
+include(${FASTDEPLOY_INSTALL_DIR}/FastDeploy.cmake)
+
+include_directories(${FASTDEPLOY_INCS})
+
+add_executable(benchmark_yolov5 ${PROJECT_SOURCE_DIR}/benchmark_yolov5.cc)
+
+if(UNIX AND (NOT APPLE) AND (NOT ANDROID))
+  target_link_libraries(benchmark_yolov5 ${FASTDEPLOY_LIBS} gflags pthread)
+else()
+  target_link_libraries(benchmark_yolov5 ${FASTDEPLOY_LIBS} gflags)
+endif()
diff --git a/benchmark/cpp/benchmark_yolov5.cc b/benchmark/cpp/benchmark_yolov5.cc
new file mode 100755
index 000000000..d84292536
--- /dev/null
+++ b/benchmark/cpp/benchmark_yolov5.cc
@@ -0,0 +1,110 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// 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.
+
+#include "fastdeploy/benchmark/utils.h"
+#include "fastdeploy/vision.h"
+#include "flags.h"
+
+bool RunModel(std::string model_file, std::string image_file, size_t warmup,
+              size_t repeats, size_t dump_period, std::string cpu_mem_file_name,
+              std::string gpu_mem_file_name) {
+  // Initialization
+  auto option = fastdeploy::RuntimeOption();
+  if (!CreateRuntimeOption(&option)) {
+    PrintUsage();
+    return false;
+  }
+  if (FLAGS_profile_mode == "runtime") {
+    option.EnableProfiling(FLAGS_include_h2d_d2h, repeats, warmup);
+  }
+  auto model = fastdeploy::vision::detection::YOLOv5(model_file, "", option);
+  if (!model.Initialized()) {
+    std::cerr << "Failed to initialize." << std::endl;
+    return false;
+  }
+  auto im = cv::imread(image_file);
+  // For Runtime
+  if (FLAGS_profile_mode == "runtime") {
+    fastdeploy::vision::DetectionResult res;
+    if (!model.Predict(im, &res)) {
+      std::cerr << "Failed to predict." << std::endl;
+      return false;
+    }
+    double profile_time = model.GetProfileTime() * 1000;
+    std::cout << "Runtime(ms): " << profile_time << "ms." << std::endl;
+    auto vis_im = fastdeploy::vision::VisDetection(im, res);
+    cv::imwrite("vis_result.jpg", vis_im);
+    std::cout << "Visualized result saved in ./vis_result.jpg" << std::endl;
+  } else {
+    // For End2End
+    // Step1: warm up for warmup times
+    std::cout << "Warmup " << warmup << " times..." << std::endl;
+    for (int i = 0; i < warmup; i++) {
+      fastdeploy::vision::DetectionResult res;
+      if (!model.Predict(im, &res)) {
+        std::cerr << "Failed to predict." << std::endl;
+        return false;
+      }
+    }
+    std::vector<float> end2end_statis;
+    // Step2: repeat for repeats times
+    std::cout << "Counting time..." << std::endl;
+    fastdeploy::TimeCounter tc;
+    fastdeploy::vision::DetectionResult res;
+    for (int i = 0; i < repeats; i++) {
+      if (FLAGS_collect_memory_info && i % dump_period == 0) {
+        fastdeploy::benchmark::DumpCurrentCpuMemoryUsage(cpu_mem_file_name);
+        fastdeploy::benchmark::DumpCurrentGpuMemoryUsage(gpu_mem_file_name,
+                                                         FLAGS_device_id);
+      }
+      tc.Start();
+      if (!model.Predict(im, &res)) {
+        std::cerr << "Failed to predict." << std::endl;
+        return false;
+      }
+      tc.End();
+      end2end_statis.push_back(tc.Duration() * 1000);
+    }
+    float end2end = std::accumulate(end2end_statis.end() - repeats,
+                                    end2end_statis.end(), 0.f) /
+                    repeats;
+    std::cout << "End2End(ms): " << end2end << "ms." << std::endl;
+    auto vis_im = fastdeploy::vision::VisDetection(im, res);
+    cv::imwrite("vis_result.jpg", vis_im);
+    std::cout << "Visualized result saved in ./vis_result.jpg" << std::endl;
+  }
+
+  return true;
+}
+
+int main(int argc, char* argv[]) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  int repeats = FLAGS_repeat;
+  int warmup = FLAGS_warmup;
+  int dump_period = FLAGS_dump_period;
+  std::string cpu_mem_file_name = "result_cpu.txt";
+  std::string gpu_mem_file_name = "result_gpu.txt";
+  // Run model
+  if (RunModel(FLAGS_model, FLAGS_image, warmup, repeats, dump_period,
+               cpu_mem_file_name, gpu_mem_file_name) != true) {
+    exit(1);
+  }
+  if (FLAGS_collect_memory_info) {
+    float cpu_mem = fastdeploy::benchmark::GetCpuMemoryUsage(cpu_mem_file_name);
+    float gpu_mem = fastdeploy::benchmark::GetGpuMemoryUsage(gpu_mem_file_name);
+    std::cout << "cpu_pss_mb: " << cpu_mem << "MB." << std::endl;
+    std::cout << "gpu_pss_mb: " << gpu_mem << "MB." << std::endl;
+  }
+  return 0;
+}
\ No newline at end of file
diff --git a/benchmark/cpp/flags.h b/benchmark/cpp/flags.h
new file mode 100755
index 000000000..3d35eb313
--- /dev/null
+++ b/benchmark/cpp/flags.h
@@ -0,0 +1,99 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// 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 "gflags/gflags.h"
+#include "fastdeploy/utils/perf.h"
+
+DEFINE_string(model, "", "Directory of the inference model.");
+DEFINE_string(image, "", "Path of the image file.");
+DEFINE_string(device, "cpu",
+              "Type of inference device, support 'cpu' or 'gpu'.");
+DEFINE_int32(device_id, 0, "device(gpu) id.");
+DEFINE_int32(warmup, 200, "Number of warmup for profiling.");
+DEFINE_int32(repeat, 1000, "Number of repeats for profiling.");
+DEFINE_string(profile_mode, "runtime", "runtime or end2end.");
+DEFINE_string(backend, "default",
+              "The inference runtime backend, support: ['default', 'ort', "
+              "'paddle', 'ov', 'trt', 'paddle_trt']");
+DEFINE_int32(cpu_thread_nums, 8, "Set numbers of cpu thread.");
+DEFINE_bool(
+    include_h2d_d2h, false, "Whether run profiling with h2d and d2h.");
+DEFINE_bool(
+    use_fp16, false,
+    "Whether to use FP16 mode, only support 'trt' and 'paddle_trt' backend");
+DEFINE_bool(
+    collect_memory_info, false, "Whether to collect memory info");
+DEFINE_int32(dump_period, 100, "How often to collect memory info.");
+
+void PrintUsage() {
+  std::cout << "Usage: infer_demo --model model_path --image img_path --device "
+               "[cpu|gpu] --backend "
+               "[default|ort|paddle|ov|trt|paddle_trt] "
+               "--use_fp16 false"
+            << std::endl;
+  std::cout << "Default value of device: cpu" << std::endl;
+  std::cout << "Default value of backend: default" << std::endl;
+  std::cout << "Default value of use_fp16: false" << std::endl;
+}
+
+bool CreateRuntimeOption(fastdeploy::RuntimeOption* option) {
+  if (FLAGS_device == "gpu") {
+    option->UseGpu();
+    if (FLAGS_backend == "ort") {
+      option->UseOrtBackend();
+    } else if (FLAGS_backend == "paddle") {
+      option->UsePaddleInferBackend();
+    } else if (FLAGS_backend == "trt" || FLAGS_backend == "paddle_trt") {
+      option->UseTrtBackend();
+      option->SetTrtInputShape("input", {1, 3, 112, 112});
+      if (FLAGS_backend == "paddle_trt") {
+        option->EnablePaddleToTrt();
+      }
+      if (FLAGS_use_fp16) {
+        option->EnableTrtFP16();
+      }
+    } else if (FLAGS_backend == "default") {
+      return true;
+    } else {
+      std::cout << "While inference with GPU, only support "
+                   "default/ort/paddle/trt/paddle_trt now, "
+                << FLAGS_backend << " is not supported." << std::endl;
+      return false;
+    }
+  } else if (FLAGS_device == "cpu") {
+    option->SetCpuThreadNum(FLAGS_cpu_thread_nums);
+    if (FLAGS_backend == "ort") {
+      option->UseOrtBackend();
+    } else if (FLAGS_backend == "ov") {
+      option->UseOpenVINOBackend();
+    } else if (FLAGS_backend == "paddle") {
+      option->UsePaddleInferBackend();
+    } else if (FLAGS_backend == "default") {
+      return true;
+    } else {
+      std::cout << "While inference with CPU, only support "
+                   "default/ort/ov/paddle now, "
+                << FLAGS_backend << " is not supported." << std::endl;
+      return false;
+    }
+  } else {
+    std::cerr << "Only support device CPU/GPU now, " << FLAGS_device
+              << " is not supported." << std::endl;
+    return false;
+  }
+
+  return true;
+}
diff --git a/benchmark/README.md b/benchmark/python/README.md
similarity index 100%
rename from benchmark/README.md
rename to benchmark/python/README.md
diff --git a/benchmark/benchmark_ernie_seq_cls.py b/benchmark/python/benchmark_ernie_seq_cls.py
similarity index 100%
rename from benchmark/benchmark_ernie_seq_cls.py
rename to benchmark/python/benchmark_ernie_seq_cls.py
diff --git a/benchmark/benchmark_ppcls.py b/benchmark/python/benchmark_ppcls.py
similarity index 96%
rename from benchmark/benchmark_ppcls.py
rename to benchmark/python/benchmark_ppcls.py
index a8219b028..20a62c9fc 100755
--- a/benchmark/benchmark_ppcls.py
+++ b/benchmark/python/benchmark_ppcls.py
@@ -17,7 +17,8 @@ import cv2
 import os
 import numpy as np
 import time
-from tqdm import tqdm 
+from tqdm import tqdm
+
 
 def parse_arguments():
     import argparse
@@ -38,19 +39,19 @@ def parse_arguments():
         "--profile_mode",
         type=str,
         default="runtime",
-        help="runtime or end2end.")      
+        help="runtime or end2end.")
     parser.add_argument(
         "--repeat",
         required=True,
         type=int,
         default=1000,
-        help="number of repeats for profiling.")    
+        help="number of repeats for profiling.")
     parser.add_argument(
         "--warmup",
         required=True,
         type=int,
         default=50,
-        help="number of warmup for profiling.")      
+        help="number of warmup for profiling.")
     parser.add_argument(
         "--device",
         default="cpu",
@@ -74,7 +75,7 @@ def parse_arguments():
         "--include_h2d_d2h",
         type=ast.literal_eval,
         default=False,
-        help="whether run profiling with h2d and d2h")       
+        help="whether run profiling with h2d and d2h")
     args = parser.parse_args()
     return args
 
@@ -85,7 +86,7 @@ def build_option(args):
     backend = args.backend
     enable_trt_fp16 = args.enable_trt_fp16
     if args.profile_mode == "runtime":
-        option.enable_profiling(args.include_h2d_d2h, args.repeat, args.warmup)    
+        option.enable_profiling(args.include_h2d_d2h, args.repeat, args.warmup)
     option.set_cpu_thread_num(args.cpu_num_thread)
     if device == "gpu":
         option.use_gpu()
@@ -274,25 +275,27 @@ if __name__ == '__main__':
             enable_gpu = args.device == "gpu"
             monitor = Monitor(enable_gpu, gpu_id)
             monitor.start()
-        
+
         im_ori = cv2.imread(args.image)
         if args.profile_mode == "runtime":
             result = model.predict(im_ori)
             profile_time = model.get_profile_time()
             dump_result["runtime"] = profile_time * 1000
-            f.writelines("Runtime(ms): {} \n".format(str(dump_result["runtime"])))
+            f.writelines("Runtime(ms): {} \n".format(
+                str(dump_result["runtime"])))
             print("Runtime(ms): {} \n".format(str(dump_result["runtime"])))
         else:
             # end2end
             for i in range(args.warmup):
                 result = model.predict(im_ori)
-            
+
             start = time.time()
             for i in tqdm(range(args.repeat)):
                 result = model.predict(im_ori)
             end = time.time()
             dump_result["end2end"] = ((end - start) / args.repeat) * 1000.0
-            f.writelines("End2End(ms): {} \n".format(str(dump_result["end2end"])))
+            f.writelines("End2End(ms): {} \n".format(
+                str(dump_result["end2end"])))
             print("End2End(ms): {} \n".format(str(dump_result["end2end"])))
 
         if enable_collect_memory_info:
@@ -304,7 +307,7 @@ if __name__ == '__main__':
                 'memory.used'] if 'gpu' in mem_info else 0
             dump_result["gpu_util"] = mem_info['gpu'][
                 'utilization.gpu'] if 'gpu' in mem_info else 0
-        
+
         if enable_collect_memory_info:
             f.writelines("cpu_rss_mb: {} \n".format(
                 str(dump_result["cpu_rss_mb"])))
diff --git a/benchmark/benchmark_ppdet.py b/benchmark/python/benchmark_ppdet.py
similarity index 95%
rename from benchmark/benchmark_ppdet.py
rename to benchmark/python/benchmark_ppdet.py
index eb0b0f4a8..c2b1da6b1 100755
--- a/benchmark/benchmark_ppdet.py
+++ b/benchmark/python/benchmark_ppdet.py
@@ -20,6 +20,7 @@ import time
 from sympy import EX
 from tqdm import tqdm
 
+
 def parse_arguments():
     import argparse
     import ast
@@ -39,19 +40,19 @@ def parse_arguments():
         "--profile_mode",
         type=str,
         default="runtime",
-        help="runtime or end2end.")      
+        help="runtime or end2end.")
     parser.add_argument(
         "--repeat",
         required=True,
         type=int,
         default=1000,
-        help="number of repeats for profiling.")    
+        help="number of repeats for profiling.")
     parser.add_argument(
         "--warmup",
         required=True,
         type=int,
         default=50,
-        help="number of warmup for profiling.")      
+        help="number of warmup for profiling.")
     parser.add_argument(
         "--device",
         default="cpu",
@@ -70,7 +71,7 @@ def parse_arguments():
         "--enable_lite_fp16",
         type=ast.literal_eval,
         default=False,
-        help="whether enable fp16 in Paddle Lite backend")    
+        help="whether enable fp16 in Paddle Lite backend")
     parser.add_argument(
         "--enable_collect_memory_info",
         type=ast.literal_eval,
@@ -80,7 +81,7 @@ def parse_arguments():
         "--include_h2d_d2h",
         type=ast.literal_eval,
         default=False,
-        help="whether run profiling with h2d and d2h")       
+        help="whether run profiling with h2d and d2h")
     args = parser.parse_args()
     return args
 
@@ -92,7 +93,7 @@ def build_option(args):
     enable_trt_fp16 = args.enable_trt_fp16
     enable_lite_fp16 = args.enable_lite_fp16
     if args.profile_mode == "runtime":
-        option.enable_profiling(args.include_h2d_d2h, args.repeat, args.warmup)  
+        option.enable_profiling(args.include_h2d_d2h, args.repeat, args.warmup)
     option.set_cpu_thread_num(args.cpu_num_thread)
     if device == "gpu":
         option.use_gpu()
@@ -149,7 +150,7 @@ def build_option(args):
         else:
             raise Exception(
                 "While inference with CPU, only support default/ort/lite/paddle now, {} is not supported.".
-                format(backend))    
+                format(backend))
     elif device == "ascend":
         option.use_ascend()
         if backend == "lite":
@@ -161,11 +162,11 @@ def build_option(args):
         else:
             raise Exception(
                 "While inference with CPU, only support default/lite now, {} is not supported.".
-                format(backend))                
+                format(backend))
     else:
         raise Exception(
-            "Only support device CPU/GPU/Kunlunxin/Ascend now, {} is not supported.".format(
-                device))
+            "Only support device CPU/GPU/Kunlunxin/Ascend now, {} is not supported.".
+            format(device))
 
     return option
 
@@ -340,19 +341,21 @@ if __name__ == '__main__':
             result = model.predict(im_ori)
             profile_time = model.get_profile_time()
             dump_result["runtime"] = profile_time * 1000
-            f.writelines("Runtime(ms): {} \n".format(str(dump_result["runtime"])))
+            f.writelines("Runtime(ms): {} \n".format(
+                str(dump_result["runtime"])))
             print("Runtime(ms): {} \n".format(str(dump_result["runtime"])))
         else:
             # end2end
             for i in range(args.warmup):
                 result = model.predict(im_ori)
-            
+
             start = time.time()
             for i in tqdm(range(args.repeat)):
                 result = model.predict(im_ori)
             end = time.time()
             dump_result["end2end"] = ((end - start) / args.repeat) * 1000.0
-            f.writelines("End2End(ms): {} \n".format(str(dump_result["end2end"])))
+            f.writelines("End2End(ms): {} \n".format(
+                str(dump_result["end2end"])))
             print("End2End(ms): {} \n".format(str(dump_result["end2end"])))
 
         if enable_collect_memory_info:
@@ -364,7 +367,7 @@ if __name__ == '__main__':
                 'memory.used'] if 'gpu' in mem_info else 0
             dump_result["gpu_util"] = mem_info['gpu'][
                 'utilization.gpu'] if 'gpu' in mem_info else 0
-        
+
         if enable_collect_memory_info:
             f.writelines("cpu_rss_mb: {} \n".format(
                 str(dump_result["cpu_rss_mb"])))
diff --git a/benchmark/benchmark_ppocr.py b/benchmark/python/benchmark_ppocr.py
similarity index 100%
rename from benchmark/benchmark_ppocr.py
rename to benchmark/python/benchmark_ppocr.py
diff --git a/benchmark/benchmark_ppseg.py b/benchmark/python/benchmark_ppseg.py
similarity index 100%
rename from benchmark/benchmark_ppseg.py
rename to benchmark/python/benchmark_ppseg.py
diff --git a/benchmark/benchmark_uie.py b/benchmark/python/benchmark_uie.py
similarity index 100%
rename from benchmark/benchmark_uie.py
rename to benchmark/python/benchmark_uie.py
diff --git a/benchmark/benchmark_yolo.py b/benchmark/python/benchmark_yolo.py
similarity index 100%
rename from benchmark/benchmark_yolo.py
rename to benchmark/python/benchmark_yolo.py
diff --git a/benchmark/convert_info.py b/benchmark/python/convert_info.py
similarity index 100%
rename from benchmark/convert_info.py
rename to benchmark/python/convert_info.py
diff --git a/benchmark/requirements.txt b/benchmark/python/requirements.txt
similarity index 100%
rename from benchmark/requirements.txt
rename to benchmark/python/requirements.txt
diff --git a/benchmark/run_benchmark_ernie_seq_cls.sh b/benchmark/python/run_benchmark_ernie_seq_cls.sh
similarity index 100%
rename from benchmark/run_benchmark_ernie_seq_cls.sh
rename to benchmark/python/run_benchmark_ernie_seq_cls.sh
diff --git a/benchmark/run_benchmark_ppcls.sh b/benchmark/python/run_benchmark_ppcls.sh
similarity index 100%
rename from benchmark/run_benchmark_ppcls.sh
rename to benchmark/python/run_benchmark_ppcls.sh
diff --git a/benchmark/run_benchmark_ppdet.sh b/benchmark/python/run_benchmark_ppdet.sh
similarity index 100%
rename from benchmark/run_benchmark_ppdet.sh
rename to benchmark/python/run_benchmark_ppdet.sh
diff --git a/benchmark/run_benchmark_ppocr.sh b/benchmark/python/run_benchmark_ppocr.sh
similarity index 100%
rename from benchmark/run_benchmark_ppocr.sh
rename to benchmark/python/run_benchmark_ppocr.sh
diff --git a/benchmark/run_benchmark_ppseg.sh b/benchmark/python/run_benchmark_ppseg.sh
similarity index 100%
rename from benchmark/run_benchmark_ppseg.sh
rename to benchmark/python/run_benchmark_ppseg.sh
diff --git a/benchmark/run_benchmark_uie.sh b/benchmark/python/run_benchmark_uie.sh
similarity index 100%
rename from benchmark/run_benchmark_uie.sh
rename to benchmark/python/run_benchmark_uie.sh
diff --git a/benchmark/run_benchmark_yolo.sh b/benchmark/python/run_benchmark_yolo.sh
similarity index 100%
rename from benchmark/run_benchmark_yolo.sh
rename to benchmark/python/run_benchmark_yolo.sh
diff --git a/fastdeploy/benchmark/benchmark.h b/fastdeploy/benchmark/benchmark.h
old mode 100644
new mode 100755
index 825fc4f54..b7463d3e9
--- a/fastdeploy/benchmark/benchmark.h
+++ b/fastdeploy/benchmark/benchmark.h
@@ -18,7 +18,7 @@
 #include "fastdeploy/benchmark/option.h"
 #include "fastdeploy/benchmark/results.h"
 
-#ifdef ENABLE_BENCHMARK                            
+#ifdef ENABLE_BENCHMARK
   #define __RUNTIME_PROFILE_LOOP_BEGIN(option, base_loop)               \
     int __p_loop = (base_loop);                                         \
     const bool __p_enable_profile = option.enable_profile;              \
@@ -75,12 +75,12 @@
         result.time_of_runtime =                                        \
           __p_tc_duration_h / static_cast<double>(__p_repeats_h);       \
       }                                                                 \
-    }  
+    }
 #else
   #define __RUNTIME_PROFILE_LOOP_BEGIN(option, base_loop)               \
-    for (int __p_i = 0; __p_i < (base_loop); ++ __p_i) { 
+    for (int __p_i = 0; __p_i < (base_loop); ++__p_i) {
   #define __RUNTIME_PROFILE_LOOP_END(result) }
   #define __RUNTIME_PROFILE_LOOP_H2D_D2H_BEGIN(option, base_loop)       \
-    for (int __p_i_h = 0; __p_i_h < (base_loop); ++ __p_i_h) {
+    for (int __p_i_h = 0; __p_i_h < (base_loop); ++__p_i_h) {
   #define __RUNTIME_PROFILE_LOOP_H2D_D2H_END(result) }
-#endif
\ No newline at end of file
+#endif
diff --git a/fastdeploy/benchmark/option.h b/fastdeploy/benchmark/option.h
old mode 100644
new mode 100755
index 6df9b473c..5af9f1585
--- a/fastdeploy/benchmark/option.h
+++ b/fastdeploy/benchmark/option.h
@@ -26,22 +26,22 @@ struct BenchmarkOption {
   int warmup = 50;              ///< Warmup for backend inference.
   int repeats = 100;            ///< Repeats for backend inference.
   bool enable_profile = false;  ///< Whether to use profile or not.
-  bool include_h2d_d2h = false; ///< Whether to include time of H2D_D2H for time of runtime.
+  bool include_h2d_d2h = false;  ///< Whether to include time of H2D_D2H for time of runtime. // NOLINT
 
   friend std::ostream& operator<<(
     std::ostream& output, const BenchmarkOption &option) {
-    if (!option.include_h2d_d2h) {                                    
-      output << "Running profiling for Runtime "                     
-             << "without H2D and D2H, ";                
-    } else {                                                             
-      output << "Running profiling for Runtime "                        
-             << "with H2D and D2H, ";                   
-    }          
-    output << "Repeats: " << option.repeats << ", "                  
-           << "Warmup: " << option.warmup;                   
-    return output;         
+    if (!option.include_h2d_d2h) {
+      output << "Running profiling for Runtime "
+             << "without H2D and D2H, ";
+    } else {
+      output << "Running profiling for Runtime "
+             << "with H2D and D2H, ";
+    }
+    output << "Repeats: " << option.repeats << ", "
+           << "Warmup: " << option.warmup;
+    return output;
   }
 };
 
-} // namespace benchmark
-} // namespace fastdeploy
\ No newline at end of file
+}  // namespace benchmark
+}  // namespace fastdeploy
diff --git a/fastdeploy/benchmark/utils.cc b/fastdeploy/benchmark/utils.cc
new file mode 100755
index 000000000..2b0bd9df1
--- /dev/null
+++ b/fastdeploy/benchmark/utils.cc
@@ -0,0 +1,122 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// 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.
+
+#include <sys/types.h>
+#if defined(__linux__) || defined(__ANDROID__)
+#include <unistd.h>
+#endif
+#include <cmath>
+
+#include "fastdeploy/benchmark/utils.h"
+
+namespace fastdeploy {
+namespace benchmark {
+
+// Remove the ch characters at both ends of str
+static std::string strip(const std::string& str, char ch = ' ') {
+  int i = 0;
+  while (str[i] == ch) {
+    i++;
+  }
+  int j = str.size() - 1;
+  while (str[j] == ch) {
+    j--;
+  }
+  return str.substr(i, j + 1 - i);
+}
+
+void DumpCurrentCpuMemoryUsage(const std::string& name) {
+#if defined(__linux__) || defined(__ANDROID__)
+  int iPid = static_cast<int>(getpid());
+  std::string command = "pmap -x " + std::to_string(iPid) + " | grep total";
+  FILE* pp = popen(command.data(), "r");
+  if (!pp) return;
+  char tmp[1024];
+
+  while (fgets(tmp, sizeof(tmp), pp) != NULL) {
+    std::ofstream write;
+    write.open(name, std::ios::app);
+    write << tmp;
+    write.close();
+  }
+  pclose(pp);
+#else
+  FDASSERT(false,
+           "Currently collect cpu memory info only supports Linux and ANDROID.")
+#endif
+  return;
+}
+
+void DumpCurrentGpuMemoryUsage(const std::string& name, int device_id) {
+#if defined(__linux__) && defined(WITH_GPU)
+  std::string command = "nvidia-smi --id=" + std::to_string(device_id) +
+                        " --query-gpu=index,uuid,name,timestamp,memory.total,"
+                        "memory.free,memory.used,utilization.gpu,utilization."
+                        "memory --format=csv,noheader,nounits";
+  FILE* pp = popen(command.data(), "r");
+  if (!pp) return;
+  char tmp[1024];
+
+  while (fgets(tmp, sizeof(tmp), pp) != NULL) {
+    std::ofstream write;
+    write.open(name, std::ios::app);
+    write << tmp;
+    write.close();
+  }
+  pclose(pp);
+#else
+  FDASSERT(false,
+           "Currently collect gpu memory info only supports Linux in GPU.")
+#endif
+  return;
+}
+
+float GetCpuMemoryUsage(const std::string& name) {
+  std::ifstream read(name);
+  std::string line;
+  float max_cpu_mem = -1;
+  while (getline(read, line)) {
+    std::stringstream ss(line);
+    std::string tmp;
+    std::vector<std::string> nums;
+    while (getline(ss, tmp, ' ')) {
+      tmp = strip(tmp);
+      if (tmp.empty()) continue;
+      nums.push_back(tmp);
+    }
+    max_cpu_mem = std::max(max_cpu_mem, stof(nums[3]));
+  }
+  return max_cpu_mem / 1024;
+}
+
+float GetGpuMemoryUsage(const std::string& name) {
+  std::ifstream read(name);
+  std::string line;
+  float max_gpu_mem = -1;
+  while (getline(read, line)) {
+    std::stringstream ss(line);
+    std::string tmp;
+    std::vector<std::string> nums;
+    while (getline(ss, tmp, ',')) {
+      tmp = strip(tmp);
+      if (tmp.empty()) continue;
+      nums.push_back(tmp);
+    }
+    max_gpu_mem = std::max(max_gpu_mem, stof(nums[6]));
+  }
+  return max_gpu_mem;
+}
+
+}  // namespace benchmark
+}  // namespace fastdeploy
diff --git a/fastdeploy/benchmark/utils.h b/fastdeploy/benchmark/utils.h
new file mode 100755
index 000000000..12770f365
--- /dev/null
+++ b/fastdeploy/benchmark/utils.h
@@ -0,0 +1,35 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// 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 "fastdeploy/utils/utils.h"
+
+namespace fastdeploy {
+namespace benchmark {
+
+// Record current cpu memory usage into file
+FASTDEPLOY_DECL void DumpCurrentCpuMemoryUsage(const std::string& name);
+
+// Record current gpu memory usage into file
+FASTDEPLOY_DECL void DumpCurrentGpuMemoryUsage(const std::string& name,
+                                               int device_id);
+
+// Get Max cpu memory usage
+FASTDEPLOY_DECL float GetCpuMemoryUsage(const std::string& name);
+
+// Get Max gpu memory usage
+FASTDEPLOY_DECL float GetGpuMemoryUsage(const std::string& name);
+
+}  // namespace benchmark
+}  // namespace fastdeploy