print average duration for caffe2 benchmark

Summary: print average duration for caffe2 benchmark

Test Plan:
buck run //xplat/caffe2:caffe2_benchmarkAppleMac -- --init_net ~/track_init_net.pb --net ~/track_predict_net.pb --warmup 10 --input 'data' --input_dims '1,4,128,256' --input_type float --iter 20
Using additional configuration options from .buckconfig.local
Building: finished in 0.6 sec (100%) 247/2137 jobs, 0 updated
  Total time: 0.6 sec
Average Duration: 18111 us

Reviewed By: larryliu0820

Differential Revision: D27745416

fbshipit-source-id: a5d20b8ef0ba4a9547d396738d5ddd1aca57684d

binaries/benchmark_helper.cc

@@ -281,13 +281,18 @@ void runNetwork(
       iter,
       ".");
   LOG(INFO) << "net runs.";
+  long long duration_sum = 0;
   for (int i = 0; i < iter; ++i) {
     caffe2::ObserverConfig::initSampleRate(1, 1, 1, 0, warmup);
     fillInputBlob(workspace, tensor_protos_map, i);
     if (wipe_cache) {
       caffe2::wipe_cache();
     }
+    auto start = std::chrono::high_resolution_clock::now();
     CAFFE_ENFORCE(net->Run(), "Main run ", i, " has failed.");
+    auto stop = std::chrono::high_resolution_clock::now();
+    auto duration = std::chrono::duration_cast<std::chrono::microseconds>(stop - start);
+    duration_sum += duration.count();
     // Write the output for the first num_blobs times
     writeOutput(
         workspace,
@@ -305,6 +310,7 @@ void runNetwork(
           std::chrono::seconds(sleep_between_iteration));
     }
   }
+  std::cout << "Average Duration: " << (duration_sum/iter) << " us" << std::endl;
   if (run_individual) {
     LOG(INFO) << "operator runs.";
     if (sleep_between_net_and_operator > 0) {