diff --git a/mace/core/runtime/hexagon/hexagon_device.cc b/mace/core/runtime/hexagon/hexagon_device.cc
index e8a0aa4e49873a306dcfa795a2a96d351f534dc5..c465c1a920930b9ef582692e4c64efc6ef5923ae 100644
--- a/mace/core/runtime/hexagon/hexagon_device.cc
+++ b/mace/core/runtime/hexagon/hexagon_device.cc
@@ -32,16 +32,18 @@ HexagonDevice::HexagonDevice(DeviceType device_type,
 
 #ifdef MACE_ENABLE_OPENCL
 GPURuntime *HexagonDevice::gpu_runtime() {
+  MACE_CHECK_NOTNULL(gpu_device_);
   return gpu_device_->gpu_runtime();
 }
 #endif  // MACE_ENABLE_OPENCL
 
 Allocator *HexagonDevice::allocator() {
 #ifdef MACE_ENABLE_OPENCL
-  return gpu_device_->allocator();
-#else
-  return allocator_.get();
+  if (gpu_device_ != nullptr) {
+    return gpu_device_->allocator();
+  }
 #endif  // MACE_ENABLE_OPENCL
+  return allocator_.get();
 }
 
 DeviceType HexagonDevice::device_type() const {
diff --git a/mace/core/runtime/hexagon/hexagon_device.h b/mace/core/runtime/hexagon/hexagon_device.h
index 660965eb9f89c46a395b0087ea6f864304a2f668..9913c893bb09c13525bf9486e989a7837d610da0 100644
--- a/mace/core/runtime/hexagon/hexagon_device.h
+++ b/mace/core/runtime/hexagon/hexagon_device.h
@@ -39,7 +39,7 @@ class HexagonDevice : public CPUDevice {
   HexagonDevice(DeviceType device_type,
 #ifdef MACE_ENABLE_OPENCL
                 utils::ThreadPool *thread_pool,
-                std::unique_ptr<GPUDevice> gpu_device);
+                std::unique_ptr<GPUDevice> gpu_device = nullptr);
 #else
                 utils::ThreadPool *thread_pool);
 #endif  // MACE_ENABLE_OPENCL
diff --git a/mace/libmace/mace.cc b/mace/libmace/mace.cc
index 44e025d09703600e315753bc0e71fd6940da066b..3085b6556de96c58eea100f8daa097420e366f9c 100644
--- a/mace/libmace/mace.cc
+++ b/mace/libmace/mace.cc
@@ -578,8 +578,9 @@ MaceEngine::Impl::Impl(const MaceEngineConfig &config)
   }
 #endif
 #if defined(MACE_ENABLE_HEXAGON) || defined(MACE_ENABLE_HTA)
-  if (device_type_ == DeviceType::HEXAGON
-      || device_type_ == DeviceType::HTA) {
+  if (device_type_ == DeviceType::HEXAGON) {
+    device_.reset(new HexagonDevice(device_type_, thread_pool_.get()));
+  } else if (device_type_ == DeviceType::HTA) {
 #ifdef MACE_ENABLE_OPENCL
     device_.reset(new HexagonDevice(
         device_type_, thread_pool_.get(),
diff --git a/mace/ops/eltwise.cc b/mace/ops/eltwise.cc
index 39f7e6487eb203fc0c0dfde869d2e661f77cd24c..94986d8481a04ea5db327dd777fdacf7383b5746 100644
--- a/mace/ops/eltwise.cc
+++ b/mace/ops/eltwise.cc
@@ -22,6 +22,7 @@
 #include <cmath>
 #include <functional>
 #include <memory>
+#include <set>
 #include <utility>
 #include <vector>
 
@@ -1198,6 +1199,27 @@ void RegisterEltwise(OpRegistry *op_registry) {
 #endif  // MACE_ENABLE_QUANTIZE
 
   MACE_REGISTER_GPU_OP(op_registry, "Eltwise", EltwiseOp);
+  MACE_REGISTER_OP_CONDITION(
+      op_registry, OpConditionBuilder("Eltwise").SetDevicePlacerFunc(
+                       [](OpConditionContext *context) -> std::set<DeviceType> {
+                         auto op = context->operator_def();
+                         if (op->output_shape_size() != op->output_size()) {
+                           return {DeviceType::CPU, DeviceType::GPU};
+                         }
+
+                         int input_size = op->input_size();
+                         auto ws = context->workspace();
+                         for (int i = 0; i < input_size; ++i) {
+                           if (ws->HasTensor(op->input(i)) &&
+                               ws->GetTensor(op->input(i))->is_weight()) {
+                             int dims = ws->GetTensor(op->input(i))->dim_size();
+                             if (dims != 1 && dims != 4) {
+                               return {DeviceType::CPU};
+                             }
+                           }
+                         }
+                         return {DeviceType::CPU, DeviceType::GPU};
+                       }));
 }
 
 }  // namespace ops
diff --git a/tools/device.py b/tools/device.py
index d1d80afe53cf1949b88281d07f4c9ce07df80be1..6ab0135bc64ae6c23359b1c89c3a375ea60da7d1 100644
--- a/tools/device.py
+++ b/tools/device.py
@@ -279,8 +279,7 @@ class DeviceWrapper:
                 if os.path.exists(opencl_parameter_file):
                     self.push(opencl_parameter_file, self.data_dir)
 
-            if self.system == SystemType.android \
-                    and device_type == common.DeviceType.HEXAGON:
+            if self.system == SystemType.android:
                 self.push(
                     "third_party/nnlib/%s/libhexagon_controller.so" % abi,
                     self.data_dir)