diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index 0e554491015218acb0593c8c8562d29c29c68bc4..d3cbcd4ce6cccc0703c95ac6bb17b8a84f1f2cf8 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -1,6 +1,7 @@
 stages:
   - cpplint
   - pycodestyle
+  - platform_compitable_tests
   - ops_test
   - ops_benchmark
   - extra_tests
@@ -34,3 +35,10 @@ extra_tests:
     - if [ -z "$TARGET_SOCS" ]; then TARGET_SOCS=random; fi
     - python tools/bazel_adb_run.py --target="//mace/kernels:kernels_test" --run_target=True --stdout_processor=unittest_stdout_processor --target_abis=armeabi-v7a,arm64-v8a --target_socs=$TARGET_SOCS
     - python tools/bazel_adb_run.py --target="//mace/utils:tuner_test" --run_target=True --stdout_processor=unittest_stdout_processor --target_abis=armeabi-v7a,arm64-v8a --target_socs=$TARGET_SOCS
+
+platform_compitable_tests:
+  stage: platform_compitable_tests
+  script:
+    - bazel build mace/core:core
+
+ 
diff --git a/mace/core/BUILD b/mace/core/BUILD
index 1b4cdaa8d51c054a9f8b55d5c7e4266273724f5d..41a3ec2d9e4e3b3286e26ff7d1ebfdaf22a24379 100644
--- a/mace/core/BUILD
+++ b/mace/core/BUILD
@@ -23,36 +23,41 @@ cc_library(
         [
             "*.cc",
             "runtime/cpu/*.cc",
-            "runtime/opencl/*.cc",
-            "runtime/hexagon/*.cc",
         ],
         exclude = [
             "*_test.cc",
-            "runtime/hexagon/hexagon_controller_dummy.cc",
         ],
-    ) + if_not_hexagon_enabled([
-        "runtime/hexagon/hexagon_controller_dummy.cc",
-    ]) + if_hexagon_enabled([
+    ) + if_android(glob(
+        [
+            "runtime/opencl/*.cc",
+        ],
+    )) + if_hexagon_enabled(glob([
+        "runtime/hexagon/*.cc",
         "runtime/hexagon/libhexagon_controller.so",
-    ]),
+    ])),
     hdrs = glob([
         "*.h",
         "runtime/cpu/*.h",
-        "runtime/opencl/*.h",
-        "runtime/hexagon/*.h",
-    ]),
-    copts = if_openmp_enabled(["-fopenmp"]),
+    ]) + if_android(glob(
+        [
+            "runtime/opencl/*.h",
+        ],
+    )) + if_hexagon_enabled(glob(["runtime/hexagon/*.h"])),
+    copts = if_openmp_enabled(["-fopenmp"]) + if_android([
+        "-DMACE_ENABLE_OPENCL",
+    ]) + if_hexagon_enabled(["-DMACE_ENABLE_HEXAGON"]),
     linkopts = ["-ldl"] + if_android([
         "-pie",
         "-lm",
     ]),
     deps = [
+        "//mace/utils",
+    ] + if_android([
         ":opencl_headers",
         "//mace/codegen:generated_opencl",
         "//mace/codegen:generated_version",
-        "//mace/utils",
         "@half//:half",
-    ] + if_production_mode([
+    ]) + if_production_mode([
         "//mace/codegen:generated_tuning_params",
         "//mace/utils:utils_prod",
     ]) + if_not_production_mode([
diff --git a/mace/core/allocator.cc b/mace/core/allocator.cc
index 567887301a6bb7ea5b5c16738fad79c454b91a60..53fd5851f22fe5d625e1121803c3713b6bee7892 100644
--- a/mace/core/allocator.cc
+++ b/mace/core/allocator.cc
@@ -13,7 +13,9 @@
 // limitations under the License.
 
 #include "mace/core/allocator.h"
+#ifdef MACE_ENABLE_OPENCL
 #include "mace/core/runtime/opencl/opencl_allocator.h"
+#endif
 
 namespace mace {
 
@@ -33,7 +35,9 @@ Allocator *GetDeviceAllocator(DeviceType type) {
 
 MACE_REGISTER_ALLOCATOR(DeviceType::CPU, new CPUAllocator());
 MACE_REGISTER_ALLOCATOR(DeviceType::NEON, new CPUAllocator());
+#ifdef MACE_ENABLE_OPENCL
 MACE_REGISTER_ALLOCATOR(DeviceType::OPENCL, new OpenCLAllocator());
+#endif
 MACE_REGISTER_ALLOCATOR(DeviceType::HEXAGON, new CPUAllocator());
 
 }  // namespace mace
diff --git a/mace/core/allocator.h b/mace/core/allocator.h
index 80466a9c1b79d936fe2c239e2872539fd828201c..c22ea47c738d551ce586c3fecb169f1f68a18e4d 100644
--- a/mace/core/allocator.h
+++ b/mace/core/allocator.h
@@ -15,10 +15,11 @@
 #ifndef MACE_CORE_ALLOCATOR_H_
 #define MACE_CORE_ALLOCATOR_H_
 
-#include <malloc.h>
+#include <stdlib.h>
 #include <map>
 #include <limits>
 #include <vector>
+#include <cstring>
 
 #include "mace/core/registry.h"
 #include "mace/core/types.h"
@@ -26,7 +27,9 @@
 
 namespace mace {
 
-#ifdef __ANDROID__
+#if defined(__hexagon__)
+constexpr size_t kMaceAlignment = 128;
+#elif defined(__ANDROID__)
 // 16 bytes = 128 bits = 32 * 4 (Neon)
 constexpr size_t kMaceAlignment = 16;
 #else
@@ -67,7 +70,7 @@ class CPUAllocator : public Allocator {
   void *New(size_t nbytes) const override {
     VLOG(3) << "Allocate CPU buffer: " << nbytes;
     void *data = nullptr;
-#ifdef __ANDROID__
+#if defined(__ANDROID__) || defined(__hexagon__)
     data = memalign(kMaceAlignment, nbytes);
 #else
     MACE_CHECK(posix_memalign(&data, kMaceAlignment, nbytes) == 0);
diff --git a/mace/core/buffer.h b/mace/core/buffer.h
index d209d52f4fea0b31370d9074705a8c865f000276..3c30e8d430bd637fd9489b7a2330a711c32c0f37 100644
--- a/mace/core/buffer.h
+++ b/mace/core/buffer.h
@@ -17,6 +17,7 @@
 
 #include <vector>
 #include <algorithm>
+#include <numeric>
 #include <functional>
 
 #include "mace/core/allocator.h"
diff --git a/mace/core/file_storage.cc b/mace/core/file_storage.cc
index 9a3f812be64a6860e144e34aec00ef9ccaad2051..5f6f07774dc12b71e259343c93f2d5b3ac603d3e 100644
--- a/mace/core/file_storage.cc
+++ b/mace/core/file_storage.cc
@@ -23,7 +23,7 @@
 #include <algorithm>
 #include <memory>
 #include <utility>
-
+#include <cstring>
 
 #include "mace/utils/logging.h"
 
diff --git a/mace/core/future.h b/mace/core/future.h
index 47c461bf03b9ec4e27d4256b764455f26e4a17d5..4dc778ad5da83613dd033bd18159086c7bed39be 100644
--- a/mace/core/future.h
+++ b/mace/core/future.h
@@ -21,7 +21,7 @@
 
 namespace mace {
 
-class CallStats;
+struct CallStats;
 
 // Wait the call to finish and get the stats if param is not nullptr
 struct StatsFuture {
diff --git a/mace/core/mace.cc b/mace/core/mace.cc
index 27d676d9b6f8cfeddb4e153f9fd82f46014f027b..76e243b138616b5dffc3cac8c7072a6bf3e18000 100644
--- a/mace/core/mace.cc
+++ b/mace/core/mace.cc
@@ -14,17 +14,23 @@
 
 #include <memory>
 
-#include "mace/core/file_storage.h"
 #include "mace/core/net.h"
-#include "mace/core/runtime/hexagon/hexagon_control_wrapper.h"
-#include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/core/types.h"
 #include "mace/public/mace.h"
 
+#ifdef MACE_ENABLE_OPENCL
+#include "mace/core/runtime/opencl/opencl_runtime.h"
+#endif  // MACE_ENABLE_OPENCL
+
+#ifdef MACE_ENABLE_HEXAGON
+#include "mace/core/runtime/hexagon/hexagon_control_wrapper.h"
+#endif  // MACE_ENABLE_HEXAGON
+
 namespace mace {
 
 // Mace Tensor
-struct MaceTensor::Impl {
+class MaceTensor::Impl {
+ public:
   std::vector<int64_t> shape;
   std::shared_ptr<float> data;
 };
@@ -49,8 +55,8 @@ MaceTensor::MaceTensor(const MaceTensor &other) {
 
 MaceTensor::MaceTensor(const MaceTensor &&other) {
   impl_ = std::unique_ptr<MaceTensor::Impl>(new MaceTensor::Impl());
-  impl_->shape = std::move(other.shape());
-  impl_->data = std::move(other.data());
+  impl_->shape = other.shape();
+  impl_->data = other.data();
 }
 
 MaceTensor &MaceTensor::operator=(const MaceTensor &other) {
@@ -60,8 +66,8 @@ MaceTensor &MaceTensor::operator=(const MaceTensor &other) {
 }
 
 MaceTensor &MaceTensor::operator=(const MaceTensor &&other) {
-  impl_->shape = std::move(other.shape());
-  impl_->data = std::move(other.data());
+  impl_->shape = other.shape();
+  impl_->data = other.data();
   return *this;
 }
 
@@ -91,7 +97,9 @@ class MaceEngine::Impl {
   DeviceType device_type_;
   std::unique_ptr<Workspace> ws_;
   std::unique_ptr<NetBase> net_;
+#ifdef MACE_ENABLE_HEXAGON
   std::unique_ptr<HexagonControlWrapper> hexagon_controller_;
+#endif
 
   DISABLE_COPY_AND_ASSIGN(Impl);
 };
@@ -103,8 +111,11 @@ MaceEngine::Impl::Impl(const NetDef *net_def,
     : op_registry_(new OperatorRegistry()),
       device_type_(device_type),
       ws_(new Workspace()),
-      net_(nullptr),
-      hexagon_controller_(nullptr) {
+      net_(nullptr)
+#ifdef MACE_ENABLE_HEXAGON
+      , hexagon_controller_(nullptr)
+#endif
+{
   LOG(INFO) << "MACE version: " << MaceVersion();
   // Set storage path for internal usage
   for (auto input_name : input_nodes) {
@@ -115,6 +126,7 @@ MaceEngine::Impl::Impl(const NetDef *net_def,
     ws_->CreateTensor(MakeString("mace_output_node_", output_name, ":0"),
                       GetDeviceAllocator(device_type_), DT_FLOAT);
   }
+#ifdef MACE_ENABLE_HEXAGON
   if (device_type == HEXAGON) {
     hexagon_controller_.reset(new HexagonControlWrapper());
     MACE_CHECK(hexagon_controller_->Config(), "hexagon config error");
@@ -130,6 +142,7 @@ MaceEngine::Impl::Impl(const NetDef *net_def,
       hexagon_controller_->PrintGraph();
     }
   } else {
+#endif
     ws_->LoadModelTensor(*net_def, device_type);
 
     // Init model
@@ -138,11 +151,14 @@ MaceEngine::Impl::Impl(const NetDef *net_def,
     if (!net->Run()) {
       LOG(FATAL) << "Net init run failed";
     }
-    net_ = std::move(CreateNet(op_registry_, *net_def, ws_.get(), device_type));
+    net_ = CreateNet(op_registry_, *net_def, ws_.get(), device_type);
+#ifdef MACE_ENABLE_HEXAGON
   }
+#endif
 }
 
 MaceEngine::Impl::~Impl() {
+#ifdef MACE_ENABLE_HEXAGON
   if (device_type_ == HEXAGON) {
     if (VLOG_IS_ON(2)) {
       hexagon_controller_->GetPerfInfo();
@@ -151,6 +167,7 @@ MaceEngine::Impl::~Impl() {
     MACE_CHECK(hexagon_controller_->TeardownGraph(), "hexagon teardown error");
     MACE_CHECK(hexagon_controller_->Finalize(), "hexagon finalize error");
   }
+#endif
 }
 
 MaceStatus MaceEngine::Impl::Run(
@@ -177,18 +194,25 @@ MaceStatus MaceEngine::Impl::Run(
         ws_->GetTensor(MakeString("mace_output_node_", output.first + ":0"));
     output_tensors.push_back(output_tensor);
   }
+#ifdef MACE_ENABLE_HEXAGON
   if (device_type_ == HEXAGON) {
     MACE_CHECK(input_tensors.size() == 1 && output_tensors.size() == 1,
                "HEXAGON not support multiple inputs and outputs yet.");
     hexagon_controller_->ExecuteGraph(*input_tensors[0], output_tensors[0]);
   } else {
+#endif
     if (!net_->Run(run_metadata)) {
       LOG(FATAL) << "Net run failed";
     }
+#ifdef MACE_ENABLE_HEXAGON
   }
+#endif
+
+#ifdef MACE_ENABLE_OPENCL
   if (device_type_ == OPENCL) {
     OpenCLRuntime::Global()->SaveBuiltCLProgram();
   }
+#endif
   for (auto &output : *outputs) {
     Tensor *output_tensor =
         ws_->GetTensor(MakeString("mace_output_node_", output.first + ":0"));
diff --git a/mace/core/mace_runtime.cc b/mace/core/mace_runtime.cc
index 4ce55861c6acba40e0a6f441c37ff4ca239d9477..8fe2b6c3b5222b7a9f68229349db040a58ea6b4d 100644
--- a/mace/core/mace_runtime.cc
+++ b/mace/core/mace_runtime.cc
@@ -13,38 +13,15 @@
 // limitations under the License.
 
 #include "mace/public/mace_runtime.h"
-#include "mace/core/runtime/cpu/cpu_runtime.h"
-#include "mace/core/runtime/opencl/opencl_runtime.h"
+#include "mace/utils/logging.h"
 
 namespace mace {
 
 std::shared_ptr<KVStorageFactory> kStorageFactory = nullptr;
 
-void SetGPUHints(GPUPerfHint gpu_perf_hint, GPUPriorityHint gpu_priority_hint) {
-  VLOG(1) << "Set GPU configurations, gpu_perf_hint: " << gpu_perf_hint
-          << ", gpu_priority_hint: " << gpu_priority_hint;
-  OpenCLRuntime::Configure(gpu_perf_hint, gpu_priority_hint);
-}
-
 void SetKVStorageFactory(std::shared_ptr<KVStorageFactory> storage_factory) {
   VLOG(1) << "Set internal KV Storage Engine";
   kStorageFactory = storage_factory;
 }
 
-MaceStatus SetOpenMPThreadPolicy(int num_threads_hint,
-                                 CPUAffinityPolicy policy) {
-  VLOG(1) << "Set CPU openmp num_threads_hint: " << num_threads_hint
-          << ", affinity policy: " << policy;
-  return SetOpenMPThreadsAndAffinityPolicy(num_threads_hint, policy);
-}
-
-void SetOpenMPThreadAffinity(int num_threads, const std::vector<int> &cpu_ids) {
-  return SetOpenMPThreadsAndAffinityCPUs(num_threads, cpu_ids);
-}
-
-MaceStatus GetBigLittleCoreIDs(std::vector<int> *big_core_ids,
-                               std::vector<int> *little_core_ids) {
-  return GetCPUBigLittleCoreIDs(big_core_ids, little_core_ids);
-}
-
 };  // namespace mace
diff --git a/mace/core/mace_types.cc b/mace/core/mace_types.cc
index c356116197e20035871a3b4e7b8fac31605f6cae..a98b1d5e7a93efaa3b36474ae7e087ac91d6cf43 100644
--- a/mace/core/mace_types.cc
+++ b/mace/core/mace_types.cc
@@ -152,12 +152,12 @@ void OperatorDef::CopyFrom(const OperatorDef &from) {
   std::copy(from_output.begin(), from_output.end(), output_.begin());
   auto from_arg = from.arg();
   arg_.resize(from_arg.size());
-  for (int i = 0; i < from_arg.size(); ++i) {
+  for (size_t i = 0; i < from_arg.size(); ++i) {
     arg_[i].CopyFrom(from_arg[i]);
   }
   auto from_output_shape = from.output_shape();
   output_shape_.resize(from_output_shape.size());
-  for (int i = 0; i < from_output_shape.size(); ++i) {
+  for (size_t i = 0; i < from_output_shape.size(); ++i) {
     output_shape_[i].CopyFrom(from_output_shape[i]);
   }
   auto from_data_type = from.output_type();
@@ -174,7 +174,7 @@ void OperatorDef::CopyFrom(const OperatorDef &from) {
   padding_ = from.padding();
   auto from_node_input = from.node_input();
   node_input_.resize(from_node_input.size());
-  for (int i = 0; i < from_node_input.size(); ++i) {
+  for (size_t i = 0; i < from_node_input.size(); ++i) {
     node_input_[i].CopyFrom(from_node_input[i]);
   }
   auto from_out_max_byte_size = from.out_max_byte_size();
@@ -223,7 +223,7 @@ void OperatorDef::add_out_max_byte_size(int value) {
 }
 const std::vector<std::string> &OperatorDef::input() const { return input_; }
 const std::string &OperatorDef::input(int index) const {
-  MACE_CHECK(0 <= index && index <= input_.size());
+  MACE_CHECK(0 <= index && index <= static_cast<int>(input_.size()));
   return input_[index];
 }
 std::string *OperatorDef::add_input() {
@@ -240,7 +240,7 @@ void OperatorDef::set_input(const std::vector<std::string> &value) {
 }
 const std::vector<std::string> &OperatorDef::output() const { return output_; }
 const std::string &OperatorDef::output(int index) const {
-  MACE_CHECK(0 <= index && index <= output_.size());
+  MACE_CHECK(0 <= index && index <= static_cast<int>(output_.size()));
   return output_[index];
 }
 std::string *OperatorDef::add_output() {
diff --git a/mace/core/net.cc b/mace/core/net.cc
index 44c889b1e1480b1d4b50001bb3e4f1dc8b7c2835..72f186c53d297db51f2a246be40b646a0ba47c23 100644
--- a/mace/core/net.cc
+++ b/mace/core/net.cc
@@ -25,7 +25,7 @@ NetBase::NetBase(const std::shared_ptr<const OperatorRegistry> op_registry,
                  const std::shared_ptr<const NetDef> net_def,
                  Workspace *ws,
                  DeviceType type)
-    : op_registry_(op_registry), name_(net_def->name()) {}
+    : name_(net_def->name()), op_registry_(op_registry) {}
 
 SerialNet::SerialNet(const std::shared_ptr<const OperatorRegistry> op_registry,
                      const std::shared_ptr<const NetDef> net_def,
diff --git a/mace/core/runtime/cpu/cpu_runtime.cc b/mace/core/runtime/cpu/cpu_runtime.cc
index 4c4e58ff035c41d2921d6c9f58a29b5f90bf2f07..f55115d6b36a22b90a0fdc3db1f8431a15642345 100644
--- a/mace/core/runtime/cpu/cpu_runtime.cc
+++ b/mace/core/runtime/cpu/cpu_runtime.cc
@@ -16,12 +16,15 @@
 
 #include <omp.h>
 #include <unistd.h>
+#include <sys/types.h>
 #include <algorithm>
 #include <utility>
 #include <vector>
 
 #include "mace/public/mace.h"
+#include "mace/public/mace_runtime.h"
 #include "mace/utils/logging.h"
+
 namespace mace {
 
 namespace {
@@ -80,8 +83,11 @@ void SortCPUIdsByMaxFreqAsc(std::vector<int> *cpu_ids, int *big_core_offset) {
 }
 
 void SetThreadAffinity(cpu_set_t mask) {
-  int sys_call_res;
+#if defined(__ANDROID__)
   pid_t pid = gettid();
+#else
+  pid_t pid = pthread_self();
+#endif
   int err = sched_setaffinity(pid, sizeof(mask), &mask);
   MACE_CHECK(err == 0, "set affinity error: ", errno);
 }
@@ -173,12 +179,28 @@ MaceStatus SetOpenMPThreadsAndAffinityPolicy(int omp_num_threads_hint,
   }
 
   if (omp_num_threads_hint <= 0 ||
-      omp_num_threads_hint > use_cpu_ids.size()) {
+      omp_num_threads_hint > static_cast<int>(use_cpu_ids.size())) {
     omp_num_threads_hint = use_cpu_ids.size();
   }
   SetOpenMPThreadsAndAffinityCPUs(omp_num_threads_hint, use_cpu_ids);
   return MACE_SUCCESS;
 }
 
+MaceStatus SetOpenMPThreadPolicy(int num_threads_hint,
+                                 CPUAffinityPolicy policy) {
+  VLOG(1) << "Set CPU openmp num_threads_hint: " << num_threads_hint
+          << ", affinity policy: " << policy;
+  return SetOpenMPThreadsAndAffinityPolicy(num_threads_hint, policy);
+}
+
+void SetOpenMPThreadAffinity(int num_threads, const std::vector<int> &cpu_ids) {
+  return SetOpenMPThreadsAndAffinityCPUs(num_threads, cpu_ids);
+}
+
+MaceStatus GetBigLittleCoreIDs(std::vector<int> *big_core_ids,
+                               std::vector<int> *little_core_ids) {
+  return GetCPUBigLittleCoreIDs(big_core_ids, little_core_ids);
+}
+
 }  // namespace mace
 
diff --git a/mace/core/runtime/opencl/opencl_runtime.cc b/mace/core/runtime/opencl/opencl_runtime.cc
index 4314edbf394b713b02d7bb70daf779c36c0b0245..baec40b2dc9c3690f4e237788e8a8c081e8fa475 100644
--- a/mace/core/runtime/opencl/opencl_runtime.cc
+++ b/mace/core/runtime/opencl/opencl_runtime.cc
@@ -22,6 +22,7 @@
 #include <vector>
 #include <utility>
 
+#include "mace/public/mace_runtime.h"
 #include "mace/core/file_storage.h"
 #include "mace/core/runtime/opencl/opencl_extension.h"
 #include "mace/public/mace.h"
@@ -35,6 +36,12 @@ extern const std::string kCompiledProgramPlatform;
 extern const std::map<std::string, std::vector<unsigned char>>
     kEncryptedProgramMap;
 
+void SetGPUHints(GPUPerfHint gpu_perf_hint, GPUPriorityHint gpu_priority_hint) {
+  VLOG(1) << "Set GPU configurations, gpu_perf_hint: " << gpu_perf_hint
+          << ", gpu_priority_hint: " << gpu_priority_hint;
+  OpenCLRuntime::Configure(gpu_perf_hint, gpu_priority_hint);
+}
+
 const std::string OpenCLErrorToString(cl_int error) {
   switch (error) {
     case CL_SUCCESS:
diff --git a/mace/core/tensor.h b/mace/core/tensor.h
index fa383aa58e75fc0846d9e7eca5ad5a9feebaa6c7..0f0b2827cbca9948402e2413942b93f629e2c85d 100644
--- a/mace/core/tensor.h
+++ b/mace/core/tensor.h
@@ -21,8 +21,10 @@
 
 #include "mace/core/buffer.h"
 #include "mace/core/preallocated_pooled_allocator.h"
-#include "mace/core/runtime/opencl/cl2_header.h"
 #include "mace/core/types.h"
+#ifdef MACE_ENABLE_OPENCL
+#include "mace/core/runtime/opencl/cl2_header.h"
+#endif
 #include "mace/public/mace.h"
 #include "mace/utils/logging.h"
 
@@ -36,6 +38,7 @@ namespace mace {
     break;                            \
   }
 
+#ifdef MACE_ENABLE_OPENCL
 #define CASES_WITH_DEFAULT(TYPE_ENUM, STMTS, INVALID, DEFAULT) \
   switch (TYPE_ENUM) {                                         \
     CASE(half, SINGLE_ARG(STMTS))                              \
@@ -56,6 +59,27 @@ namespace mace {
       DEFAULT;                                                 \
       break;                                                   \
   }
+#else
+#define CASES_WITH_DEFAULT(TYPE_ENUM, STMTS, INVALID, DEFAULT) \
+  switch (TYPE_ENUM) {                                         \
+    CASE(float, SINGLE_ARG(STMTS))                             \
+    CASE(double, SINGLE_ARG(STMTS))                            \
+    CASE(int32_t, SINGLE_ARG(STMTS))                           \
+    CASE(uint8_t, SINGLE_ARG(STMTS))                           \
+    CASE(uint16_t, SINGLE_ARG(STMTS))                          \
+    CASE(int16_t, SINGLE_ARG(STMTS))                           \
+    CASE(int8_t, SINGLE_ARG(STMTS))                            \
+    CASE(std::string, SINGLE_ARG(STMTS))                       \
+    CASE(int64_t, SINGLE_ARG(STMTS))                           \
+    CASE(bool, SINGLE_ARG(STMTS))                              \
+    case DT_INVALID:                                           \
+      INVALID;                                                 \
+      break;                                                   \
+    default:                                                   \
+      DEFAULT;                                                 \
+      break;                                                   \
+  }
+#endif
 
 #define CASES(TYPE_ENUM, STMTS)                                      \
   CASES_WITH_DEFAULT(TYPE_ENUM, STMTS, LOG(FATAL) << "Type not set"; \
@@ -137,6 +161,7 @@ class Tensor {
     return buffer_ != nullptr && !buffer_->OnHost() && !has_opencl_image();
   }
 
+#ifdef MACE_ENABLE_OPENCL
   inline cl::Image *opencl_image() const {
     MACE_CHECK(has_opencl_image(), "do not have image");
     return static_cast<cl::Image *>(buffer_->buffer());
@@ -146,6 +171,7 @@ class Tensor {
     MACE_CHECK(has_opencl_buffer(), "do not have opencl buffer");
     return static_cast<cl::Buffer *>(buffer_->buffer());
   }
+#endif
 
   inline index_t buffer_offset() const { return buffer_->offset(); }
 
diff --git a/mace/core/types.cc b/mace/core/types.cc
index 9bab2dab2f8e5cbe8e5cf7f47bff699ff22a5587..44fd1d9271c1e5e62d94a7ff2de10e5322568124 100644
--- a/mace/core/types.cc
+++ b/mace/core/types.cc
@@ -40,11 +40,18 @@ bool DataTypeCanUseMemcpy(DataType dt) {
 
 std::string DataTypeToString(const DataType dt) {
   static std::map<DataType, std::string> dtype_string_map = {
-      {DT_FLOAT, "DT_FLOAT"},   {DT_HALF, "DT_HALF"},
-      {DT_DOUBLE, "DT_DOUBLE"}, {DT_UINT8, "DT_UINT8"},
-      {DT_INT8, "DT_INT8"},     {DT_INT32, "DT_INT32"},
-      {DT_UINT32, "DT_UINT32"}, {DT_UINT16, "DT_UINT16"},
-      {DT_INT64, "DT_INT64"},   {DT_BOOL, "DT_BOOL"},
+      {DT_FLOAT, "DT_FLOAT"},
+#ifdef MACE_ENABLE_OPENCL
+      {DT_HALF, "DT_HALF"},
+#endif
+      {DT_DOUBLE, "DT_DOUBLE"},
+      {DT_UINT8, "DT_UINT8"},
+      {DT_INT8, "DT_INT8"},
+      {DT_INT32, "DT_INT32"},
+      {DT_UINT32, "DT_UINT32"},
+      {DT_UINT16, "DT_UINT16"},
+      {DT_INT64, "DT_INT64"},
+      {DT_BOOL, "DT_BOOL"},
       {DT_STRING, "DT_STRING"}};
   MACE_CHECK(dt != DT_INVALID) << "Not support Invalid data type";
   return dtype_string_map[dt];
@@ -54,8 +61,10 @@ size_t GetEnumTypeSize(const DataType dt) {
   switch (dt) {
     case DT_FLOAT:
       return sizeof(float);
+#ifdef MACE_ENABLE_OPENCL
     case DT_HALF:
       return sizeof(half);
+#endif
     case DT_UINT8:
       return sizeof(uint8_t);
     case DT_INT8:
diff --git a/mace/core/types.h b/mace/core/types.h
index 7b63fa290c076fba23063ce13207f0b5429ceae6..c07c48297f046bb1f3624c36c523ab7c7797353b 100644
--- a/mace/core/types.h
+++ b/mace/core/types.h
@@ -19,13 +19,17 @@
 #include <string>
 
 #include "mace/public/mace_types.h"
+#ifdef MACE_ENABLE_OPENCL
 #include "include/half.hpp"
+#endif
 
 namespace mace {
 
 typedef int64_t index_t;
 
+#ifdef MACE_ENABLE_OPENCL
 using half = half_float::half;
+#endif
 
 bool DataTypeCanUseMemcpy(DataType dt);
 
@@ -62,7 +66,9 @@ struct EnumToDataType {};  // Specializations below
     typedef TYPE Type;                      \
   }
 
+#ifdef MACE_ENABLE_OPENCL
 MATCH_TYPE_AND_ENUM(half, DT_HALF);
+#endif
 MATCH_TYPE_AND_ENUM(float, DT_FLOAT);
 MATCH_TYPE_AND_ENUM(double, DT_DOUBLE);
 MATCH_TYPE_AND_ENUM(int32_t, DT_INT32);
diff --git a/mace/core/workspace.cc b/mace/core/workspace.cc
index 107ac8fa5e74a5258d7b7a674a676e66ebe3d0fb..0c681b14b70d2df9c81773652413b0a140513358 100644
--- a/mace/core/workspace.cc
+++ b/mace/core/workspace.cc
@@ -32,8 +32,7 @@ Tensor *Workspace::CreateTensor(const std::string &name,
     VLOG(3) << "Tensor " << name << " already exists. Skipping.";
   } else {
     VLOG(3) << "Creating Tensor " << name;
-    tensor_map_[name] =
-        std::move(std::unique_ptr<Tensor>(new Tensor(alloc, type)));
+    tensor_map_[name] = std::unique_ptr<Tensor>(new Tensor(alloc, type));
     tensor_map_[name]->SetSourceOpName(name);
   }
   return GetTensor(name);
@@ -82,12 +81,12 @@ void Workspace::LoadModelTensor(const NetDef &net_def, DeviceType type) {
   }
   VLOG(3) << "Model data size: " << model_data_size;
 
-  if (type == DeviceType::CPU) {
-    tensor_buffer_ = std::move(std::unique_ptr<Buffer>(
-        new Buffer(GetDeviceAllocator(type), model_data_ptr, model_data_size)));
+  if (type == DeviceType::CPU || type == DeviceType::NEON) {
+    tensor_buffer_ = std::unique_ptr<Buffer>(
+        new Buffer(GetDeviceAllocator(type), model_data_ptr, model_data_size));
   } else {
-    tensor_buffer_ = std::move(std::unique_ptr<Buffer>(
-        new Buffer(GetDeviceAllocator(type), model_data_size)));
+    tensor_buffer_ = std::unique_ptr<Buffer>(
+        new Buffer(GetDeviceAllocator(type), model_data_size));
     tensor_buffer_->Map(nullptr);
     tensor_buffer_->Copy(model_data_ptr, 0, model_data_size);
     tensor_buffer_->UnMap();
diff --git a/mace/kernels/BUILD b/mace/kernels/BUILD
index 1551eb47e05ba4dac9face7e2e6f783f60448817..4eb4b8508409e3c1a57a28cb4a1f198409573334 100644
--- a/mace/kernels/BUILD
+++ b/mace/kernels/BUILD
@@ -7,7 +7,7 @@ package(
 
 licenses(["notice"])  # Apache 2.0
 
-load("//mace:mace.bzl", "if_android", "if_neon_enabled", "if_openmp_enabled", "if_android_armv7")
+load("//mace:mace.bzl", "if_android", "if_neon_enabled", "if_openmp_enabled", "if_android_armv7", "if_hexagon_enabled")
 
 cc_library(
     name = "kernels",
@@ -28,7 +28,9 @@ cc_library(
         "opencl/*.h",
         "arm/*.h",
     ]),
-    copts = if_openmp_enabled(["-fopenmp"]) + if_neon_enabled(["-DMACE_ENABLE_NEON"]) + if_android_armv7(["-mfpu=neon -mfloat-abi=softfp"]),
+    copts = if_openmp_enabled(["-fopenmp"]) + if_neon_enabled(["-DMACE_ENABLE_NEON"]) + if_android_armv7(["-mfpu=neon -mfloat-abi=softfp"]) + if_android([
+        "-DMACE_ENABLE_OPENCL",
+    ]) + if_hexagon_enabled(["-DMACE_ENABLE_HEXAGON"]),
     linkopts = if_android(["-lm"]),
     deps = [
         "//mace/core",
@@ -46,7 +48,9 @@ cc_test(
             "opencl/*_test.cc",
         ],
     ),
-    copts = if_openmp_enabled(["-fopenmp"]) + if_neon_enabled(["-DMACE_ENABLE_NEON"]),
+    copts = if_openmp_enabled(["-fopenmp"]) + if_neon_enabled(["-DMACE_ENABLE_NEON"]) + if_android_armv7(["-mfpu=neon -mfloat-abi=softfp"]) + if_android([
+        "-DMACE_ENABLE_OPENCL",
+    ]) + if_hexagon_enabled(["-DMACE_ENABLE_HEXAGON"]),
     linkopts = ["-fopenmp"],
     linkstatic = 1,
     deps = [
@@ -55,15 +59,3 @@ cc_test(
         "@gtest//:gtest_main",
     ],
 )
-
-cc_test(
-    name = "benchmark",
-    testonly = 1,
-    srcs = glob(["benchmark/*.cc"]),
-    linkstatic = 1,
-    deps = [
-        ":kernels",
-        "//mace/core",
-        "//mace/core:test_benchmark_main",
-    ],
-)
diff --git a/mace/ops/BUILD b/mace/ops/BUILD
index c7a9b95b40bc8588ebf9dbbae11e6433ea26defe..131beceb222f34accb201e731205cb8b03425718 100644
--- a/mace/ops/BUILD
+++ b/mace/ops/BUILD
@@ -7,7 +7,7 @@ package(
 
 licenses(["notice"])  # Apache 2.0
 
-load("//mace:mace.bzl", "if_android", "if_neon_enabled", "if_openmp_enabled", "if_android_armv7")
+load("//mace:mace.bzl", "if_android", "if_neon_enabled", "if_openmp_enabled", "if_android_armv7", "if_hexagon_enabled")
 
 cc_library(
     name = "test",
@@ -34,7 +34,9 @@ cc_library(
         ["*.h"],
         exclude = ["ops_test_util.h"],
     ),
-    copts = if_openmp_enabled(["-fopenmp"]) + if_neon_enabled(["-DMACE_ENABLE_NEON"]) + if_android_armv7(["-mfpu=neon -mfloat-abi=softfp"]),
+    copts = if_openmp_enabled(["-fopenmp"]) + if_neon_enabled(["-DMACE_ENABLE_NEON"]) + if_android_armv7(["-mfpu=neon -mfloat-abi=softfp"]) + if_android([
+        "-DMACE_ENABLE_OPENCL",
+    ]) + if_hexagon_enabled(["-DMACE_ENABLE_HEXAGON"]),
     deps = [
         "//mace/kernels",
     ],
@@ -47,6 +49,9 @@ cc_test(
     srcs = glob(
         ["*_test.cc"],
     ),
+    copts = if_openmp_enabled(["-fopenmp"]) + if_neon_enabled(["-DMACE_ENABLE_NEON"]) + if_android_armv7(["-mfpu=neon -mfloat-abi=softfp"]) + if_android([
+        "-DMACE_ENABLE_OPENCL",
+    ]) + if_hexagon_enabled(["-DMACE_ENABLE_HEXAGON"]),
     linkopts = ["-fopenmp"],
     linkstatic = 1,
     deps = [
@@ -60,6 +65,9 @@ cc_test(
     name = "ops_benchmark",
     testonly = 1,
     srcs = glob(["*_benchmark.cc"]),
+    copts = if_openmp_enabled(["-fopenmp"]) + if_neon_enabled(["-DMACE_ENABLE_NEON"]) + if_android_armv7(["-mfpu=neon -mfloat-abi=softfp"]) + if_android([
+        "-DMACE_ENABLE_OPENCL",
+    ]) + if_hexagon_enabled(["-DMACE_ENABLE_HEXAGON"]),
     linkopts = ["-fopenmp"],
     linkstatic = 1,
     deps = [
diff --git a/mace/public/mace_types.h b/mace/public/mace_types.h
index 4a7fcc30bb58fb7e06aa2c3275142f2caebd15cb..48da1dd3bc6f9b063a09465d97ec461676882b0e 100644
--- a/mace/public/mace_types.h
+++ b/mace/public/mace_types.h
@@ -352,7 +352,8 @@ struct OperatorStats {
   CallStats stats;
 };
 
-struct RunMetadata {
+class RunMetadata {
+ public:
   std::vector<OperatorStats> op_stats;
 };
 
diff --git a/mace/utils/env_time.h b/mace/utils/env_time.h
index c670e83fe500267e18730418fdf2660f8f26d80e..d70124b585cacf174facfb94728515d2dc5f48ba 100644
--- a/mace/utils/env_time.h
+++ b/mace/utils/env_time.h
@@ -16,15 +16,22 @@
 #define MACE_UTILS_ENV_TIME_H_
 
 #include <stdint.h>
+#ifdef __hexagon__
+#include <HAP_perf.h>
+#else
 #include <sys/time.h>
-#include <time.h>
+#endif
 
 namespace mace {
 
 inline int64_t NowMicros() {
+#ifdef __hexagon__
+  return HAP_perf_get_time_us();
+#else
   struct timeval tv;
   gettimeofday(&tv, nullptr);
   return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
+#endif
 }
 
 }  // namespace mace
diff --git a/mace/utils/memory_logging.h b/mace/utils/memory_logging.h
index a2f861fad4562832c3117374294ac4a99ecbec92..7111d7301352ae2f6208905e4e22af5321624ac0 100644
--- a/mace/utils/memory_logging.h
+++ b/mace/utils/memory_logging.h
@@ -15,13 +15,16 @@
 #ifndef MACE_UTILS_MEMORY_LOGGING_H_
 #define MACE_UTILS_MEMORY_LOGGING_H_
 
+#ifndef __hexagon__
 #include <malloc.h>
+#endif
 #include <string>
 
 #include "mace/utils/logging.h"
 
 namespace mace {
 
+#ifdef MACE_ENABLE_MEMORY_LOGGING
 class MallinfoChangeLogger {
  public:
   explicit MallinfoChangeLogger(const std::string &name) : name_(name) {
@@ -97,7 +100,6 @@ class MallinfoChangeLogger {
   }
 };
 
-#ifdef MACE_ENABLE_MEMORY_LOGGING
 #define MACE_MEMORY_LOGGING_GUARD()                                        \
   MallinfoChangeLogger mem_logger_##__line__(std::string(__FILE__) + ":" + \
                                              std::string(__func__));
diff --git a/mace/utils/utils.h b/mace/utils/utils.h
index fecf03ec64ce92d6f2aad208dd1f1fa4c678d7f1..9085006061762d8304303f4258eba503947b60ec 100644
--- a/mace/utils/utils.h
+++ b/mace/utils/utils.h
@@ -15,7 +15,6 @@
 #ifndef MACE_UTILS_UTILS_H_
 #define MACE_UTILS_UTILS_H_
 
-#include <sys/time.h>
 #include <sstream>
 #include <string>
 #include <utility>
@@ -60,7 +59,7 @@ inline std::string ObfuscateString(const std::string &src,
   for (size_t i = 0; i < src.size(); i++) {
     dest[i] = src[i] ^ lookup_table[i % lookup_table.size()];
   }
-  return std::move(dest);
+  return dest;
 }
 
 // ObfuscateString(ObfuscateString(str)) ==> str
@@ -96,7 +95,7 @@ inline std::string ObfuscateSymbol(const std::string &src) {
     // There is no collision if it's true for every char at every position
     dest[i] = encode_dict[(idx + i + 31) % encode_dict.size()];
   }
-  return std::move(dest);
+  return dest;
 }
 
 #ifdef MACE_OBFUSCATE_LITERALS