diff --git a/mace/kernels/opencl/winograd_transform.cc b/mace/kernels/opencl/winograd_transform.cc
index a842ba719174505e115fd26bc8bd8cd7f2301898..f74eb191d885647a42bfef5ec81e44c01f09efed 100644
--- a/mace/kernels/opencl/winograd_transform.cc
+++ b/mace/kernels/opencl/winograd_transform.cc
@@ -54,8 +54,8 @@ void WinogradTransformFunctor<DeviceType::OPENCL, T>::operator()(const Tensor *i
     kernel_.setArg(idx++, static_cast<uint32_t>(paddings[1] / 2));
   }
 
-  const uint32_t gws[2] = {static_cast<size_t>(out_width),
-                         static_cast<size_t>(RoundUpDiv4(input_tensor->dim(3)))};
+  const uint32_t gws[2] = {static_cast<uint32_t>(out_width),
+                         static_cast<uint32_t>(RoundUpDiv4(input_tensor->dim(3)))};
   const std::vector<uint32_t> lws = {128, 8, 1};
   std::stringstream ss;
   ss << "winograd_transform_kernel_"
@@ -126,8 +126,8 @@ void WinogradInverseTransformFunctor<DeviceType::OPENCL, T>::operator()(const Te
     kernel_.setArg(idx++, prelu_alpha_);
   }
 
-  const uint32_t gws[2] = {static_cast<size_t>(input_tensor->dim(2)),
-                         static_cast<size_t>(RoundUpDiv4(input_tensor->dim(1)))};
+  const uint32_t gws[2] = {static_cast<uint32_t>(input_tensor->dim(2)),
+                         static_cast<uint32_t>(RoundUpDiv4(input_tensor->dim(1)))};
   const std::vector<uint32_t> lws = {128, 8, 1};
 
   std::stringstream ss;
diff --git a/mace/kernels/reshape.h b/mace/kernels/reshape.h
index a185567f769214dcd5398d48845de5ea5d4a6346..4d37a19974d683c3f916e2ccc170dfac11ed94c7 100644
--- a/mace/kernels/reshape.h
+++ b/mace/kernels/reshape.h
@@ -20,6 +20,7 @@ struct ReshapeFunctor {
                   Tensor *output,
                   StatsFuture *future) {
     output->Resize(out_shape);
+    // TODO copy on write to avoid this copy.
     output->CopyBytes(input->raw_data(), input->size() * sizeof(T));
   }
 };
diff --git a/mace/ops/eltwise_benchmark.cc b/mace/ops/eltwise_benchmark.cc
index 8cd222333e62576719ae688fe347dbd9fc45c072..80a3f072b0c785bf26342408b6a91f9b98b63831 100644
--- a/mace/ops/eltwise_benchmark.cc
+++ b/mace/ops/eltwise_benchmark.cc
@@ -15,8 +15,8 @@ static void EltwiseBenchmark(int iters, kernels::EltwiseType type, int n, int h,
 
   OpsTestNet net;
   // Add input data
-  net.AddRandomInput<D, float>("Input0", {n, h, w, c});
-  net.AddRandomInput<D, float>("Input1", {n, h, w, c});
+  net.AddRandomInput<D, T>("Input0", {n, h, w, c});
+  net.AddRandomInput<D, T>("Input1", {n, h, w, c});
 
   if (D == DeviceType::OPENCL) {
     BufferToImage<D, half>(net, "Input0", "InputImg0", kernels::BufferType::IN_OUT_CHANNEL);
@@ -26,7 +26,7 @@ static void EltwiseBenchmark(int iters, kernels::EltwiseType type, int n, int h,
         .Input("InputImg1")
         .AddIntArg("type", static_cast<int>(type))
         .AddFloatsArg("coeff", {1.2, 2.1})
-        .AddIntArg("T", static_cast<int>(DT_HALF))
+        .AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
         .Output("OutputImg")
         .Finalize(net.NewOperatorDef());
   } else {
@@ -63,16 +63,17 @@ static void EltwiseBenchmark(int iters, kernels::EltwiseType type, int n, int h,
   }                                                                         \
   BENCHMARK(BM_ELTWISE_##ELT_TYPE##_##N##_##H##_##W##_##C##_##TYPE##_##DEVICE)
 
-#define BM_ELTWISE(ELT_TYPE, N, H, W, C, TYPE)       \
-  BM_ELTWISE_MACRO(ELT_TYPE, N, H, W, C, TYPE, CPU); \
-  BM_ELTWISE_MACRO(ELT_TYPE, N, H, W, C, TYPE, OPENCL);
+#define BM_ELTWISE(ELT_TYPE, N, H, W, C, )       \
+  BM_ELTWISE_MACRO(ELT_TYPE, N, H, W, C, float, CPU); \
+  BM_ELTWISE_MACRO(ELT_TYPE, N, H, W, C, float, OPENCL); \
+  BM_ELTWISE_MACRO(ELT_TYPE, N, H, W, C, half, OPENCL);
 
-BM_ELTWISE(0, 1, 256, 256, 32, float);
-BM_ELTWISE(0, 1, 128, 128, 32, float);
-BM_ELTWISE(1, 1, 128, 128, 32, float);
-BM_ELTWISE(2, 1, 128, 128, 32, float);
-BM_ELTWISE(0, 1, 240, 240, 256, float);
-BM_ELTWISE(1, 1, 240, 240, 256, float);
-BM_ELTWISE(2, 1, 240, 240, 256, float);
+BM_ELTWISE(0, 1, 256, 256, 32);
+BM_ELTWISE(0, 1, 128, 128, 32);
+BM_ELTWISE(1, 1, 128, 128, 32);
+BM_ELTWISE(2, 1, 128, 128, 32);
+BM_ELTWISE(0, 1, 240, 240, 256);
+BM_ELTWISE(1, 1, 240, 240, 256);
+BM_ELTWISE(2, 1, 240, 240, 256);
 
 }  //  namespace mace
diff --git a/mace/ops/reshape.cc b/mace/ops/reshape.cc
index c6ec12d646c59c1796c3d8a6f94eb9a77ed00508..d72052713c13d76bee21d9a44291e2197b71ffd5 100644
--- a/mace/ops/reshape.cc
+++ b/mace/ops/reshape.cc
@@ -8,9 +8,9 @@ namespace mace {
 
 void Register_Reshape(OperatorRegistry *op_registry) {
   REGISTER_OPERATOR(op_registry, OpKeyBuilder("Reshape")
-      .Device(DeviceType::CPU)
-      .TypeConstraint<float>("T")
-      .Build(),
+                                     .Device(DeviceType::CPU)
+                                     .TypeConstraint<float>("T")
+                                     .Build(),
                     ReshapeOp<DeviceType::CPU, float>);
 }