diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index fe0fd433b61fdcb48acd4aa2e733f137d7919fe2..c9545fd88f94342e10c9db9f6814eceddb528d0d 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -99,7 +99,7 @@ python_tools_tests:
   script:
     - pwd
     - rm -rf mace-models
-    - GIT_SSH_COMMAND="ssh -o UserKnownHostsFile=/dev/null -o StrictHostKeyChecking=no" git clone git@v9.git.n.xiaomi.com:deep-computing/mace-models.git
+    - GIT_SSH_COMMAND="ssh -o UserKnownHostsFile=/dev/null -o StrictHostKeyChecking=no" git clone git@github.com:XiaoMi/mace-models.git
     - CONF_FILE=mace-models/mobilenet-v2/mobilenet-v2.yml
     - sh -c "python tools/converter.py build --config=${CONF_FILE} --disable_tuning && python tools/converter.py run --config=${CONF_FILE} --round=1 --validate && python tools/converter.py run --config=${CONF_FILE} --example --round=1 --validate" || exit 1
     - rm -rf mace-models
diff --git a/mace/benchmark/benchmark_model.cc b/mace/benchmark/benchmark_model.cc
index 5be6d77778963531ecebc0b0cb6aad1963732846..7d2c17f6b7b7fba760369dad4514cce038f1b105 100644
--- a/mace/benchmark/benchmark_model.cc
+++ b/mace/benchmark/benchmark_model.cc
@@ -253,9 +253,12 @@ int Main(int argc, char **argv) {
   mace::DeviceType device_type = ParseDeviceType(FLAGS_device);
 
   // config runtime
-  mace::SetOpenMPThreadPolicy(
+  MaceStatus ret = mace::SetOpenMPThreadPolicy(
       FLAGS_omp_num_threads,
       static_cast<CPUAffinityPolicy >(FLAGS_cpu_affinity_policy));
+  if (ret != MACE_SUCCESS) {
+    LOG(WARNING) << "Set openmp or cpu affinity failed.";
+  }
 #ifdef MACE_ENABLE_OPENCL
   if (device_type == DeviceType::GPU) {
     mace::SetGPUHints(
diff --git a/mace/core/mace.cc b/mace/core/mace.cc
index 2150e83afcdbf05795360bfef7e55cbe4b38b08a..db04fcc699c481f0e06c385a7b7cc8325a463aa7 100644
--- a/mace/core/mace.cc
+++ b/mace/core/mace.cc
@@ -204,9 +204,6 @@ MaceStatus MaceEngine::Impl::Init(
     MACE_CHECK(hexagon_controller_->Init(), "hexagon init error");
     hexagon_controller_->SetDebugLevel(
         static_cast<int>(mace::logging::LogMessage::MinVLogLevel()));
-    int dsp_mode =
-        ProtoArgHelper::GetOptionalArg<NetDef, int>(*net_def, "dsp_mode", 0);
-    hexagon_controller_->SetGraphMode(dsp_mode);
     MACE_CHECK(hexagon_controller_->SetupGraph(*net_def, model_data),
                "hexagon setup graph error");
     if (VLOG_IS_ON(2)) {
diff --git a/mace/core/runtime/cpu/cpu_runtime.cc b/mace/core/runtime/cpu/cpu_runtime.cc
index 1a070c8fb1bce26343ec2c202d782803b0e2ad64..ec1ed80eed340f63ac30d053c2becb3418c43e25 100644
--- a/mace/core/runtime/cpu/cpu_runtime.cc
+++ b/mace/core/runtime/cpu/cpu_runtime.cc
@@ -77,14 +77,19 @@ int GetCPUMaxFreq(int cpu_id) {
   return freq;
 }
 
-void SetThreadAffinity(cpu_set_t mask) {
+MaceStatus SetThreadAffinity(cpu_set_t mask) {
 #if defined(__ANDROID__)
   pid_t pid = gettid();
 #else
   pid_t pid = syscall(SYS_gettid);
 #endif
   int err = sched_setaffinity(pid, sizeof(mask), &mask);
-  MACE_CHECK(err == 0, "set affinity error: ", strerror(errno));
+  if (err) {
+    LOG(WARNING) << "set affinity error: " << strerror(errno);
+    return MACE_INVALID_ARGS;
+  } else {
+    return MACE_SUCCESS;
+  }
 }
 
 }  // namespace
@@ -129,8 +134,8 @@ MaceStatus GetCPUBigLittleCoreIDs(std::vector<int> *big_core_ids,
   return MACE_SUCCESS;
 }
 
-void SetOpenMPThreadsAndAffinityCPUs(int omp_num_threads,
-                                     const std::vector<int> &cpu_ids) {
+MaceStatus SetOpenMPThreadsAndAffinityCPUs(int omp_num_threads,
+                                           const std::vector<int> &cpu_ids) {
 #ifdef MACE_ENABLE_OPENMP
   VLOG(1) << "Set OpenMP threads number: " << omp_num_threads
           << ", CPU core IDs: " << MakeString(cpu_ids);
@@ -146,17 +151,23 @@ void SetOpenMPThreadsAndAffinityCPUs(int omp_num_threads,
   for (auto cpu_id : cpu_ids) {
     CPU_SET(cpu_id, &mask);
   }
-
 #ifdef MACE_ENABLE_OPENMP
+  std::vector<MaceStatus> status(omp_num_threads);
 #pragma omp parallel for
   for (int i = 0; i < omp_num_threads; ++i) {
     VLOG(1) << "Set affinity for OpenMP thread " << omp_get_thread_num()
             << "/" << omp_get_num_threads();
-    SetThreadAffinity(mask);
+    status[i] = SetThreadAffinity(mask);
+  }
+  for (int i = 0; i < omp_num_threads; ++i) {
+    if (status[i] != MACE_SUCCESS)
+      return MACE_INVALID_ARGS;
   }
+  return MACE_SUCCESS;
 #else
-  SetThreadAffinity(mask);
+  MaceStatus status = SetThreadAffinity(mask);
   VLOG(1) << "Set affinity without OpenMP: " << mask.__bits[0];
+  return status;
 #endif
 }
 
@@ -191,8 +202,8 @@ MaceStatus SetOpenMPThreadsAndAffinityPolicy(int omp_num_threads_hint,
       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;
+
+  return SetOpenMPThreadsAndAffinityCPUs(omp_num_threads_hint, use_cpu_ids);
 }
 
 MaceStatus SetOpenMPThreadPolicy(int num_threads_hint,
@@ -202,7 +213,8 @@ MaceStatus SetOpenMPThreadPolicy(int num_threads_hint,
   return SetOpenMPThreadsAndAffinityPolicy(num_threads_hint, policy);
 }
 
-void SetOpenMPThreadAffinity(int num_threads, const std::vector<int> &cpu_ids) {
+MaceStatus SetOpenMPThreadAffinity(int num_threads,
+                                   const std::vector<int> &cpu_ids) {
   return SetOpenMPThreadsAndAffinityCPUs(num_threads, cpu_ids);
 }
 
diff --git a/mace/core/runtime/cpu/cpu_runtime.h b/mace/core/runtime/cpu/cpu_runtime.h
index 6a305f5aaee9ea1d3f459ccf330d49bc0aa4ca78..333729e19491449d5d519a4ce78eae1ab3d4cef1 100644
--- a/mace/core/runtime/cpu/cpu_runtime.h
+++ b/mace/core/runtime/cpu/cpu_runtime.h
@@ -25,8 +25,8 @@ namespace mace {
 MaceStatus GetCPUBigLittleCoreIDs(std::vector<int> *big_core_ids,
                                   std::vector<int> *little_core_ids);
 
-void SetOpenMPThreadsAndAffinityCPUs(int omp_num_threads,
-                                     const std::vector<int> &cpu_ids);
+MaceStatus SetOpenMPThreadsAndAffinityCPUs(int omp_num_threads,
+                                           const std::vector<int> &cpu_ids);
 
 MaceStatus SetOpenMPThreadsAndAffinityPolicy(int omp_num_threads_hint,
                                              CPUAffinityPolicy policy);
diff --git a/mace/core/runtime/hexagon/hexagon_control_wrapper.cc b/mace/core/runtime/hexagon/hexagon_control_wrapper.cc
index 0dac0c327692fbc998c2e6f03bd930f0a93f7b09..62312744845055bc05370dc51bb6e8e7ed33c812 100644
--- a/mace/core/runtime/hexagon/hexagon_control_wrapper.cc
+++ b/mace/core/runtime/hexagon/hexagon_control_wrapper.cc
@@ -51,7 +51,7 @@ int HexagonControlWrapper::GetVersion() {
 
 bool HexagonControlWrapper::Config() {
   LOG(INFO) << "Hexagon config";
-  if (hexagon_controller_InitHexagonWithMaxAttributes(0, 100) != 0) {
+  if (hexagon_nn_set_powersave_level(0) != 0) {
     return false;
   }
   return hexagon_nn_config() == 0;
@@ -66,7 +66,7 @@ bool HexagonControlWrapper::Init() {
 
 bool HexagonControlWrapper::Finalize() {
   LOG(INFO) << "Hexagon finalize";
-  return hexagon_controller_DeInitHexagon() == 0;
+  return hexagon_nn_set_powersave_level(1) == 0;
 }
 
 bool HexagonControlWrapper::SetupGraph(const NetDef &net_def,
@@ -271,11 +271,6 @@ void HexagonControlWrapper::SetDebugLevel(int level) {
              "set debug level error");
 }
 
-void HexagonControlWrapper::SetGraphMode(int mode) {
-  LOG(INFO) << "Set dsp mode: " << mode;
-  MACE_CHECK(hexagon_nn_set_graph_mode(nn_id_, mode) == 0, "set mode error");
-}
-
 void HexagonControlWrapper::GetPerfInfo() {
   LOG(INFO) << "Get perf info";
   std::vector<hexagon_nn_perfinfo> perf_info(MACE_MAX_NODE);
diff --git a/mace/core/runtime/hexagon/hexagon_control_wrapper.h b/mace/core/runtime/hexagon/hexagon_control_wrapper.h
index 4e4d8499e45c1b174caa06c8e3c2be679affd288..adcd5224e447ae565a1286246ae84e1be23d08f8 100644
--- a/mace/core/runtime/hexagon/hexagon_control_wrapper.h
+++ b/mace/core/runtime/hexagon/hexagon_control_wrapper.h
@@ -17,10 +17,10 @@
 
 #include <vector>
 
-#include "mace/core/runtime/hexagon/hexagon_controller.h"
 #include "mace/core/runtime/hexagon/quantize.h"
 #include "mace/core/tensor.h"
 #include "mace/public/mace.h"
+#include "third_party/nnlib/hexagon_nn.h"
 
 namespace mace {
 
@@ -44,7 +44,6 @@ class HexagonControlWrapper {
   void GetPerfInfo();
   void ResetPerfInfo();
   void SetDebugLevel(int level);
-  void SetGraphMode(int mode);
 
  private:
   static constexpr int NODE_ID_OFFSET = 10000;
diff --git a/mace/core/runtime/hexagon/hexagon_controller.h b/mace/core/runtime/hexagon/hexagon_controller.h
deleted file mode 100644
index afe4b727de4936aedee3c9515963d013050e48eb..0000000000000000000000000000000000000000
--- a/mace/core/runtime/hexagon/hexagon_controller.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// Copyright 2018 Xiaomi, Inc.  All rights reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef MACE_CORE_RUNTIME_HEXAGON_HEXAGON_CONTROLLER_H_
-#define MACE_CORE_RUNTIME_HEXAGON_HEXAGON_CONTROLLER_H_
-
-#include "third_party/nnlib/hexagon_nn.h"
-
-#ifdef __cplusplus
-extern "C" {
-#else
-#include <stdbool.h>
-#endif  // __cplusplus
-
-int hexagon_controller_InitHexagonWithMaxAttributes(int enable_dcvs,
-                                                    int bus_usage);
-
-int hexagon_controller_DeInitHexagon();
-
-#ifdef __cplusplus
-}
-#endif  // __cplusplus
-
-#endif  // MACE_CORE_RUNTIME_HEXAGON_HEXAGON_CONTROLLER_H_
-
diff --git a/mace/core/testing/test_benchmark_main.cc b/mace/core/testing/test_benchmark_main.cc
index 014a2b826806ec29fccc0e527e4f83975dbfc8b5..4f68d7063af63a2bfd0c124d347a4c527b84dd44 100644
--- a/mace/core/testing/test_benchmark_main.cc
+++ b/mace/core/testing/test_benchmark_main.cc
@@ -18,6 +18,7 @@
 #include "mace/core/testing/test_benchmark.h"
 #include "mace/public/mace.h"
 #include "mace/public/mace_runtime.h"
+#include "mace/utils/logging.h"
 
 DEFINE_string(filter, "all", "op benchmark regex filter, eg:.*CONV.*");
 DEFINE_int32(gpu_perf_hint, 3, "0:DEFAULT/1:LOW/2:NORMAL/3:HIGH");
@@ -33,9 +34,12 @@ int main(int argc, char **argv) {
   gflags::ParseCommandLineFlags(&argc, &argv, true);
 
   // config runtime
-  mace::SetOpenMPThreadPolicy(
+  mace::MaceStatus status = mace::SetOpenMPThreadPolicy(
       FLAGS_omp_num_threads,
       static_cast<mace::CPUAffinityPolicy >(FLAGS_cpu_affinity_policy));
+  if (status != mace::MACE_SUCCESS) {
+    LOG(WARNING) << "Set openmp or cpu affinity failed.";
+  }
   mace::SetGPUHints(
       static_cast<mace::GPUPerfHint>(FLAGS_gpu_perf_hint),
       static_cast<mace::GPUPriorityHint>(FLAGS_gpu_priority_hint));
diff --git a/mace/examples/android/app/build.gradle b/mace/examples/android/app/build.gradle
index b9101ae015408ba68c81700254a728c98f1f5ced..0e4d948e80ac3a865536028274dcd900d095871a 100644
--- a/mace/examples/android/app/build.gradle
+++ b/mace/examples/android/app/build.gradle
@@ -4,7 +4,7 @@ android {
     compileSdkVersion 26
     defaultConfig {
         applicationId "com.xiaomi.mace.demo"
-        minSdkVersion 23
+        minSdkVersion 19
         targetSdkVersion 26
         versionCode 1
         versionName "1.0"
diff --git a/mace/examples/android/app/src/main/java/com/xiaomi/mace/demo/result/InitData.java b/mace/examples/android/app/src/main/java/com/xiaomi/mace/demo/result/InitData.java
index c65483426ef7159d01462d7f39a24488d9351aaa..ab0f54b587bc9bbcd574c12b6f449172cade1689 100644
--- a/mace/examples/android/app/src/main/java/com/xiaomi/mace/demo/result/InitData.java
+++ b/mace/examples/android/app/src/main/java/com/xiaomi/mace/demo/result/InitData.java
@@ -34,7 +34,7 @@ public class InitData {
     public InitData() {
         model = MODELS[0];
         ompNumThreads = 2;
-        cpuAffinityPolicy = 1;
+        cpuAffinityPolicy = 0;
         gpuPerfHint = 3;
         gpuPriorityHint = 3;
         device = DEVICES[0];
diff --git a/mace/examples/android/build.sh b/mace/examples/android/build.sh
index 1533e80695299f140fb769af93aae5f3ea2fcc06..c1344c82872333931700fefdaa96837ec6e09c1c 100755
--- a/mace/examples/android/build.sh
+++ b/mace/examples/android/build.sh
@@ -5,8 +5,8 @@ set -e -u -o pipefail
 pushd ../../../
 python tools/converter.py build --config=docs/getting_started/models/demo_app_models.yaml
 
-cp -r builds/mobilenet/include/ mace/examples/android/macelibrary/src/main/cpp/
-cp -r builds/mobilenet/lib/ mace/examples/android/macelibrary/src/main/cpp/
+cp -r builds/mobilenet/include mace/examples/android/macelibrary/src/main/cpp/
+cp -r builds/mobilenet/lib mace/examples/android/macelibrary/src/main/cpp/
 
 popd
 
diff --git a/mace/examples/android/macelibrary/build.gradle b/mace/examples/android/macelibrary/build.gradle
index 1ee095f45af487c85b25c378700095ed7182d173..2f25f01728428faf03a648f72b7b70ddec3b2824 100644
--- a/mace/examples/android/macelibrary/build.gradle
+++ b/mace/examples/android/macelibrary/build.gradle
@@ -9,7 +9,7 @@ android {
     }
 
     defaultConfig {
-        minSdkVersion 23
+        minSdkVersion 19
         targetSdkVersion 26
         versionCode 1
         versionName "1.0"
diff --git a/mace/public/mace_runtime.h b/mace/public/mace_runtime.h
index 7f8b0576b462d1b4c992e87c091cbfe2c0ed9300..f7fbf430dddbce6747fd3b953c9778d655c523c0 100644
--- a/mace/public/mace_runtime.h
+++ b/mace/public/mace_runtime.h
@@ -125,7 +125,8 @@ MaceStatus SetOpenMPThreadPolicy(int num_threads_hint,
 // affinity to offline cores may run very slow or unexpectedly. In such cases,
 // please use SetOpenMPThreadPolicy with default policy instead.
 __attribute__((visibility("default")))
-void SetOpenMPThreadAffinity(int num_threads, const std::vector<int> &cpu_ids);
+MaceStatus SetOpenMPThreadAffinity(int num_threads,
+                                   const std::vector<int> &cpu_ids);
 
 // Get ARM big.LITTLE configuration.
 //
diff --git a/mace/tools/validation/mace_run.cc b/mace/tools/validation/mace_run.cc
index cba3a926622d8f0867ff1e54ac0425f9752e456f..c50f7502c4abacd77cca8e32898df14f63d48fae 100644
--- a/mace/tools/validation/mace_run.cc
+++ b/mace/tools/validation/mace_run.cc
@@ -198,9 +198,12 @@ bool RunModel(const std::string &model_name,
               const std::vector<std::vector<int64_t>> &output_shapes) {
   DeviceType device_type = ParseDeviceType(FLAGS_device);
   // config runtime
-  mace::SetOpenMPThreadPolicy(
+  MaceStatus status = mace::SetOpenMPThreadPolicy(
       FLAGS_omp_num_threads,
       static_cast<CPUAffinityPolicy >(FLAGS_cpu_affinity_policy));
+  if (status != MACE_SUCCESS) {
+    LOG(WARNING) << "Set openmp or cpu affinity failed.";
+  }
 #ifdef MACE_ENABLE_OPENCL
   if (device_type == DeviceType::GPU) {
     mace::SetGPUHints(
diff --git a/third_party/nnlib/hexagon_nn.h b/third_party/nnlib/hexagon_nn.h
index b8dbad353a059c2e55c10785b33c97880ca1dadb..b7485c93f4db702b60091b66ac2a9af9b2438719 100644
--- a/third_party/nnlib/hexagon_nn.h
+++ b/third_party/nnlib/hexagon_nn.h
@@ -37,7 +37,7 @@
 #define THIRD_PARTY_NNLIB_HEXAGON_NN_H_
 #ifndef __QAIC_HEADER
 #define __QAIC_HEADER(ff) ff
-#endif  // __QAIC_HEADER
+#endif  //__QAIC_HEADER
 
 #ifndef __QAIC_HEADER_EXPORT
 #define __QAIC_HEADER_EXPORT
@@ -49,7 +49,7 @@
 
 #ifndef __QAIC_IMPL
 #define __QAIC_IMPL(ff) ff
-#endif  // __QAIC_IMPL
+#endif  //__QAIC_IMPL
 
 #ifndef __QAIC_IMPL_EXPORT
 #define __QAIC_IMPL_EXPORT
@@ -131,7 +131,7 @@ struct hexagon_nn_const_node {
 };
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_config)(void)
     __QAIC_HEADER_ATTRIBUTE;
-__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_init)(hexagon_nn_nn_id* g)
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_init)(hexagon_nn_nn_id *g)
     __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_set_debug_level)(
     hexagon_nn_nn_id id, int level) __QAIC_HEADER_ATTRIBUTE;
@@ -221,6 +221,8 @@ __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_execute_new)(
     int inputsLen,
     hexagon_nn_tensordef *outputs,
     int outputsLen) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT unsigned int __QAIC_HEADER(hexagon_nn_get_dsp_offset)(void)
+    __QAIC_HEADER_ATTRIBUTE;
 #ifdef __cplusplus
 }
 #endif
diff --git a/third_party/nnlib/libhexagon_controller.so b/third_party/nnlib/libhexagon_controller.so
index 7daedd621b12686032f952a0da6564a13b63981a..05c8f3d8d71401ff117ce9b75f21cebc5e44dc66 100755
Binary files a/third_party/nnlib/libhexagon_controller.so and b/third_party/nnlib/libhexagon_controller.so differ