Change the way to sync at batch norm opencl benchmark.

mace/ops/batch_norm_benchmark.cc

@@ -39,14 +39,14 @@ static void BatchNorm(
   // Warm-up
   for (int i = 0; i < 5; ++i) {
     net.RunOp(D);
-    net.Sync();
   }
+  net.Sync();
 
   mace::testing::StartTiming();
   while (iters--) {
     net.RunOp(D);
-    net.Sync();
   }
+  net.Sync();
 }
 
 #define BM_BATCH_NORM_MACRO(N, C, H, W, TYPE, DEVICE)                  \

mace/ops/ops_test_util.h

@@ -141,6 +141,7 @@ class OpsTestNet {
     net_def.add_op()->CopyFrom(op_def_);
     VLOG(3) << net_def.DebugString();
     net_ = CreateNet(net_def, &ws_, device);
+    device_ = device;
     return net_->Run();
   }
 
@@ -151,7 +152,7 @@ class OpsTestNet {
   }
 
   void Sync() {
-    if (net_) {
+    if (net_ && device_ == DeviceType::OPENCL) {
       OpenCLRuntime::Get()->command_queue().finish();
     }
   }
@@ -160,6 +161,7 @@ class OpsTestNet {
   Workspace ws_;
   OperatorDef op_def_;
   std::unique_ptr<NetBase> net_;
+  DeviceType device_;
 };
 
 class OpsTestBase : public ::testing::Test {