[MOBILE][OPENCL] Enhance mobilenet gpu demo. test=develop (#3156)

mobile/test/net/test_mobilenet_GPU.cpp

@@ -17,47 +17,128 @@ limitations under the License. */
 #include "../test_helper.h"
 #include "../test_include.h"
 
-int main() {
+int main(int argc, char **argv) {
+  // init input args
+  string model_dir = g_mobilenet;
+  int64_t N = 1;
+  int64_t C = 3;
+  int64_t H = 224;
+  int64_t W = 224;
+  int repeats = 10;
+  int warmup = 10;
+  int print_output_elem = 0;
+
+  std::cout << "argc:" << argc << std::endl;
+  if (argc > 1 && argc < 9) {
+    std::cout << "usage:" << argv[0] << "\n"
+              << " <model_dir>\n"
+              << " <input_n>\n"
+              << " <input_c>\n"
+              << " <input_h>\n"
+              << " <input_w>\n"
+              << " <repeats>\n"
+              << " <warmup>\n"
+              << " <print_output>"
+              << std::endl;
+    return 0;
+  }
+
+  if (argc >= 9) {
+    model_dir = argv[1];
+    N = atoi(argv[2]);
+    C = atoi(argv[3]);
+    H = atoi(argv[4]);
+    W = atoi(argv[5]);
+    repeats = atoi(argv[6]);
+    warmup = atoi(argv[7]);
+    print_output_elem = atoi(argv[8]);
+  }
+
+  std::cout << "input shape(NCHW):"
+            << N << " "
+            << C << " "
+            << H << " "
+            << W << std::endl;
+  std::cout << "repeats:" << repeats << std::endl;
+  std::cout << "model_dir:" << model_dir << std::endl;
+
   paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL> paddle_mobile;
   //    paddle_mobile.SetThreadNum(4);
-  auto time1 = paddle_mobile::time();
+  auto load_start = paddle_mobile::time();
 #ifdef PADDLE_MOBILE_CL
   paddle_mobile.SetCLPath("/data/local/tmp/bin");
 #endif
 
-  //  auto isok = paddle_mobile.Load(
-  //      std::string(g_mobilenet_vision) + "/vision_mobilenet_model",
-  //      std::string(g_mobilenet_vision) + "/vision_mobilenet_params", true);
+  auto load_model_status = paddle_mobile.Load(std::string(model_dir), true);
+  if (!load_model_status) {
+    std::cout << "failed to load model from:" << model_dir << std::endl;
+    return 0;
+  }
 
-  auto isok = paddle_mobile.Load(std::string(g_mobilenet), true);
-  if (isok) {
-    auto time2 = paddle_mobile::time();
-    std::cout << "load cost :" << paddle_mobile::time_diff(time1, time2) << "ms"
-              << std::endl;
+  auto load_end = paddle_mobile::time();
+  std::cout << "load cost:"
+            << paddle_mobile::time_diff(load_start, load_end)
+            << " ms"  << std::endl;
 
+  // input tensor
   std::vector<float> input;
-    std::vector<int64_t> dims{1, 3, 224, 224};
+  std::vector<int64_t> dims{N, C, H, W};
   GetInput<float>(g_test_image_1x3x224x224_banana, &input, dims);
 
+  // warmup
   std::vector<float> vec_result = paddle_mobile.Predict(input, dims);
+  for (int widx = 0; widx < warmup; ++widx) {
+    paddle_mobile.Predict(input, dims);
+  }
 
-    auto time3 = paddle_mobile::time();
-    int max = 1;
-    for (int i = 0; i < max; ++i) {
+  // benchmark
+  float sum_duration = 0.0f;
+  float min_duration = 1e5f;
+  float max_duration = 1e-5f;
+  float ave_duration = -1;
+  for (int ridx = 0; ridx < repeats; ++ridx) {
+    auto start = paddle_mobile::time();
     vec_result = paddle_mobile.Predict(input, dims);
+    auto end = paddle_mobile::time();
+    auto duration = paddle_mobile::time_diff(start, end);
+    sum_duration += duration;
+    min_duration = (duration > min_duration) ? min_duration : duration;
+    max_duration = (duration < max_duration) ? max_duration : duration;
+    std::cout << "ridx:" << ridx + 1 << "/" << repeats
+              << " " << duration << " ms" << std::endl;
   }
-    auto time4 = paddle_mobile::time();
 
-    std::cout << "predict cost :"
-              << paddle_mobile::time_diff(time3, time4) / max << "ms"
-              << std::endl;
+  // benchmark result
+  ave_duration = sum_duration / static_cast<float>(repeats);
+
+  // output result
+  float output_sum = 0;
+  float output_ave = -1;
+  for (size_t oidx = 0; oidx < vec_result.size(); ++oidx) {
+    output_sum += vec_result[oidx];
+    if (print_output_elem) {
+      std::cout << "out_idx:" << oidx << " " << vec_result[oidx] << std::endl;
+    }
+  }
+  output_ave = output_sum / static_cast<float>(vec_result.size());
   std::vector<float>::iterator biggest =
       std::max_element(std::begin(vec_result), std::end(vec_result));
-    std::cout << " Max element is " << *biggest << " at position "
-              << std::distance(std::begin(vec_result), biggest) << std::endl;
-  }
 
-  std::cout << "如果结果Nan请查看: test/images/g_test_image_1x3x224x224_banana "
+  // summary
+  std::cout << "===== predict benchmark ====" << std::endl
+            << "run repeats:" << repeats << std::endl
+            << "sum_duration:" << sum_duration << " ms" << std::endl
+            << "ave_duration:" << ave_duration << " ms" << std::endl
+            << "max_duration:" << max_duration << " ms" << std::endl
+            << "min_duration:" << min_duration << " ms" << std::endl
+            << "\n===== predict result ====" << std::endl
+            << "output_sum:" << output_sum << std::endl
+            << "output_ave:" << output_ave << std::endl
+            << "output_size:" << vec_result.size() << std::endl
+            << "Max element is " << *biggest << " at position "
+            << std::distance(std::begin(vec_result), biggest) << std::endl
+            << "Note: 如果结果Nan请查看:"
+               " test/images/g_test_image_1x3x224x224_banana "
                "是否存在?"
             << std::endl;
   return 0;