[framework][place] remove prefered_place and kHost in valid_places (#2192)

* [framework][place] remove prefered_place, use place order in valid_place array instead test=develop * remove kHost from valid_places test=develop

[framework][place] remove prefered_place and kHost in valid_places (#2192)
* [framework][place] remove prefered_place, use place order in valid_place array instead test=develop * remove kHost from valid_places test=develop
17833acb · sangoly · GitHub · 31ab471e · 17833acb · 17833acb
42 changed file
--- a/lite/api/android/jni/native/convert_util_jni.h
+++ b/lite/api/android/jni/native/convert_util_jni.h
@@ -145,8 +145,6 @@ inline CxxConfig jcxxconfig_to_cpp_cxxconfig(JNIEnv *env, jobject jcxxconfig) {
  jmethodID model_dir_method =
      env->GetMethodID(cxxconfig_jclazz, "getModelDir", "()Ljava/lang/String;");
-  jmethodID preferred_place_method = env->GetMethodID(
-      cxxconfig_jclazz, "getPreferredPlace", "()Lcom/baidu/paddle/lite/Place;");
  jmethodID valid_places_method = env->GetMethodID(
      cxxconfig_jclazz, "getValidPlaces", "()[Lcom/baidu/paddle/lite/Place;");
@@ -159,13 +157,6 @@ inline CxxConfig jcxxconfig_to_cpp_cxxconfig(JNIEnv *env, jobject jcxxconfig) {
    config.set_model_dir(cpp_model_dir);
  }
-  jobject java_preferred_place =
-      env->CallObjectMethod(jcxxconfig, preferred_place_method);
-  if (java_preferred_place != nullptr) {
-    Place cpp_preferred_place = jplace_to_cpp_place(env, java_preferred_place);
-    config.set_preferred_place(cpp_preferred_place);
-  }
  jobject object_valid_places =
      env->CallObjectMethod(jcxxconfig, valid_places_method);
  jobjectArray *java_valid_places =

--- a/lite/api/android/jni/src/com/baidu/paddle/lite/CxxConfig.java
+++ b/lite/api/android/jni/src/com/baidu/paddle/lite/CxxConfig.java
@@ -18,17 +18,8 @@ package com.baidu.paddle.lite;
 */
 public class CxxConfig extends ConfigBase {
-    protected Place preferredPlace;
    protected Place[] validPlaces;
-    public Place getPreferredPlace() {
-        return preferredPlace;
-    }
-    public void setPreferredPlace(Place preferredPlace) {
-        this.preferredPlace = preferredPlace;
-    }
    public Place[] getValidPlaces() {
        return validPlaces;
    }

--- a/lite/api/apis_test.cc
+++ b/lite/api/apis_test.cc
@@ -51,17 +51,12 @@ bool CompareTensors(const std::string& name,
 TEST(CXXApi_LightApi, optim_model) {
  lite::Predictor cxx_api;
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
      Place{TARGET(kX86), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},  // Both works on X86 and ARM
  });
  // On ARM devices, the preferred X86 target not works, but it can still
  // select ARM kernels.
-  cxx_api.Build(FLAGS_model_dir,
+  cxx_api.Build(FLAGS_model_dir, "", "", valid_places);
-                "",
-                "",
-                Place{TARGET(kX86), PRECISION(kFloat)},
-                valid_places);
  cxx_api.SaveModel(FLAGS_optimized_model);
 }
@@ -72,17 +67,12 @@ TEST(CXXApi_LightApi, save_and_load_model) {
  // CXXAPi
  {
    std::vector<Place> valid_places({
-        Place{TARGET(kHost), PRECISION(kFloat)},
        Place{TARGET(kX86), PRECISION(kFloat)},
        Place{TARGET(kARM), PRECISION(kFloat)},  // Both works on X86 and ARM
    });
    // On ARM devices, the preferred X86 target not works, but it can still
    // select ARM kernels.
-    cxx_api.Build(FLAGS_model_dir,
+    cxx_api.Build(FLAGS_model_dir, "", "", valid_places);
-                  "",
-                  "",
-                  Place{TARGET(kX86), PRECISION(kFloat)},
-                  valid_places);
    auto* x = cxx_api.GetInput(0);
    SetConstInput(x);

--- a/lite/api/benchmark.cc
+++ b/lite/api/benchmark.cc
@@ -47,9 +47,7 @@ void OutputOptModel(const std::string& load_model_dir,
      Place{TARGET(kARM), PRECISION(kInt8)},
      Place{TARGET(kX86), PRECISION(kFloat)},
      Place{TARGET(kOpenCL), PRECISION(kFloat)},
-      Place{TARGET(kHost), PRECISION(kFloat)},
  });
-  config.set_preferred_place(Place{TARGET(kARM), PRECISION(kFloat)});
  auto predictor = lite_api::CreatePaddlePredictor(config);
  int ret = system(

--- a/lite/api/cxx_api.cc
+++ b/lite/api/cxx_api.cc
@@ -110,14 +110,12 @@ void Predictor::Build(const lite_api::CxxConfig &config,
  const std::string &model_path = config.model_dir();
  const std::string &model_file = config.model_file();
  const std::string &param_file = config.param_file();
-  const Place prefer_place = config.preferred_place();
  const bool model_from_memory = config.model_from_memory();
  LOG(INFO) << "load from memory " << model_from_memory;
  Build(model_path,
        model_file,
        param_file,
-        prefer_place,
        valid_places,
        passes,
        model_type,
@@ -126,7 +124,6 @@ void Predictor::Build(const lite_api::CxxConfig &config,
 void Predictor::Build(const std::string &model_path,
                      const std::string &model_file,
                      const std::string &param_file,
-                      const Place &prefer_place,
                      const std::vector<Place> &valid_places,
                      const std::vector<std::string> &passes,
                      lite_api::LiteModelType model_type,
@@ -153,21 +150,24 @@ void Predictor::Build(const std::string &model_path,
    default:
      LOG(FATAL) << "Unknown model type";
  }
-  Build(program_desc_, prefer_place, valid_places, passes);
+  Build(program_desc_, valid_places, passes);
 }
 void Predictor::Build(const cpp::ProgramDesc &desc,
-                      const Place &prefer_place,
                      const std::vector<Place> &valid_places,
                      const std::vector<std::string> &passes) {
  program_desc_ = desc;
-  Program program(desc, scope_, valid_places);
+  std::vector<Place> inner_places = valid_places;
-  optimizer_.KernelPickPreferPlace(prefer_place);
+  inner_places.emplace_back(TARGET(kHost), PRECISION(kAny), DATALAYOUT(kAny));
+  inner_places.emplace_back(
+      TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW));
+  Program program(desc, scope_, inner_places);
+  /// The first place in valid_places is
  core::KernelPickFactor factor;
  factor.ConsiderTarget();
  factor.ConsiderPrecision();
  factor.ConsiderDataLayout();
-  optimizer_.Run(std::move(program), valid_places, factor, passes);
+  optimizer_.Run(std::move(program), inner_places, factor, passes);
  exec_scope_ = optimizer_.exec_scope();
 }

--- a/lite/api/cxx_api.h
+++ b/lite/api/cxx_api.h
@@ -50,14 +50,12 @@ class LITE_API Predictor {
      const std::string& model_path,
      const std::string& model_file_path,
      const std::string& param_file_path,
-      const Place& prefer_place,
      const std::vector<Place>& valid_places,
      const std::vector<std::string>& passes = {},
      lite_api::LiteModelType model_type = lite_api::LiteModelType::kProtobuf,
      bool memory_from_memory = false);
  void Build(const cpp::ProgramDesc& desc,
-             const Place& prefer_place,
             const std::vector<Place>& valid_places,
             const std::vector<std::string>& passes = {});
@@ -132,10 +130,8 @@ class LITE_API Predictor {
 class LITE_API CXXTrainer {
 public:
  CXXTrainer(const std::shared_ptr<lite::Scope>& root_scope,
-             const Place& preferred_place,
             const std::vector<Place>& valid_places)
      : scope_(root_scope),
-        preferred_place_(preferred_place),
        valid_places_(valid_places),
        main_program_executor_(Predictor(scope_)) {}
@@ -144,7 +140,7 @@ class LITE_API CXXTrainer {
  // NOTE Just support to execute the 0-th block currently.
  Predictor& BuildMainProgramExecutor(const framework::proto::ProgramDesc& desc,
                                      int block_id = 0) {
-    main_program_executor_.Build(desc, preferred_place_, valid_places_);
+    main_program_executor_.Build(desc, valid_places_);
    return main_program_executor_;
  }
@@ -162,14 +158,12 @@ class LITE_API CXXTrainer {
  void RunStartupProgram(const framework::proto::ProgramDesc& desc,
                         int block_id = 0) {
    Predictor exe(scope_);
-    exe.Build(desc, preferred_place_, valid_places_);
+    exe.Build(desc,  valid_places_);
    exe.Run();
  }
 private:
  std::shared_ptr<lite::Scope> scope_;
-  Place preferred_place_;
  std::vector<Place> valid_places_;
  // The training program.

--- a/lite/api/cxx_api_bin.cc
+++ b/lite/api/cxx_api_bin.cc
@@ -35,13 +35,11 @@ void Run(const char* model_dir, int repeat) {
 #endif
  lite::Predictor predictor;
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
-      Place{TARGET(kARM), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kInt8)},
+      Place{TARGET(kARM), PRECISION(kFloat)},
  });
-  predictor.Build(
+  predictor.Build(model_dir, "", "", valid_places);
-      model_dir, "", "", Place{TARGET(kARM), PRECISION(kInt8)}, valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));

--- a/lite/api/cxx_api_impl.cc
+++ b/lite/api/cxx_api_impl.cc
@@ -62,7 +62,6 @@ void CxxPaddleApiImpl::Init(const lite_api::CxxConfig &config) {
  Env<TARGET(kCUDA)>::Init();
 #endif
  auto places = config.valid_places();
-  places.emplace_back(TARGET(kHost), PRECISION(kAny), DATALAYOUT(kAny));
  raw_predictor_.Build(config, places);
  raw_predictor_.PrepareFeedFetch();
 }

--- a/lite/api/cxx_api_test.cc
+++ b/lite/api/cxx_api_test.cc
@@ -43,13 +43,8 @@ TEST(CXXApi, test) {
 TEST(CXXApi, save_model) {
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kX86), PRECISION(kFloat)}});
-                                   Place{TARGET(kX86), PRECISION(kFloat)}});
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
-  predictor.Build(FLAGS_model_dir,
-                  "",
-                  "",
-                  Place{TARGET(kCUDA), PRECISION(kFloat)},
-                  valid_places);
  LOG(INFO) << "Save optimized model to " << FLAGS_optimized_model;
  predictor.SaveModel(FLAGS_optimized_model,
@@ -59,11 +54,11 @@ TEST(CXXApi, save_model) {
 }
 /*TEST(CXXTrainer, train) {
-  Place prefer_place({TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW)});
+  Place place({TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW)});
-  std::vector<Place> valid_places({prefer_place});
+  std::vector<Place> valid_places({place});
  auto scope = std::make_shared<lite::Scope>();
-  CXXTrainer trainer(scope, prefer_place, valid_places);
+  CXXTrainer trainer(scope, valid_places);
  std::string main_program_pb, startup_program_pb;
  ReadBinaryFile(FLAGS_main_program_path, &main_program_pb);
@@ -94,13 +89,8 @@ TEST(CXXApi, save_model) {
 #ifdef LITE_WITH_ARM
 TEST(CXXApi, save_model) {
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kARM), PRECISION(kFloat)}});
-                                   Place{TARGET(kARM), PRECISION(kFloat)}});
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
-  predictor.Build(FLAGS_model_dir,
-                  "",
-                  "",
-                  Place{TARGET(kARM), PRECISION(kFloat)},
-                  valid_places);
  LOG(INFO) << "Save optimized model to " << FLAGS_optimized_model;
  predictor.SaveModel(FLAGS_optimized_model);
@@ -110,12 +100,10 @@ TEST(CXXApi, save_model) {
 TEST(CXXApi, load_model_naive) {
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kARM), PRECISION(kFloat)}});
-                                   Place{TARGET(kARM), PRECISION(kFloat)}});
  predictor.Build(FLAGS_optimized_model + ".naive",
                  "",
                  "",
-                  Place{TARGET(kARM), PRECISION(kFloat)},
                  valid_places,
                  {},
                  lite_api::LiteModelType::kNaiveBuffer);

--- a/lite/api/detection_model_test.cc
+++ b/lite/api/detection_model_test.cc
@@ -34,7 +34,6 @@ void OutputOptModel(const std::string& load_model_dir,
                    const std::string& save_optimized_model_dir) {
  lite_api::CxxConfig config;
  config.set_model_dir(load_model_dir);
-  config.set_preferred_place(Place{TARGET(kX86), PRECISION(kFloat)});
  config.set_valid_places({
      Place{TARGET(kX86), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},

--- a/lite/api/efficientnet_b0_test.cc
+++ b/lite/api/efficientnet_b0_test.cc
@@ -25,13 +25,12 @@
 namespace paddle {
 namespace lite {
-void TestModel(const std::vector<Place> &valid_places,
+void TestModel(const std::vector<Place> &valid_places) {
-               const Place &preferred_place) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_HIGH, FLAGS_threads);
  lite::Predictor predictor;
-  predictor.Build(FLAGS_model_dir, "", "", preferred_place, valid_places);
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
  auto *input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));
@@ -80,22 +79,20 @@ void TestModel(const std::vector<Place> &valid_places,
 TEST(EfficientNetB0, test_arm) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},
      // Place{TARGET(kOpenCL), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kARM), PRECISION(kFloat)}));
+  TestModel(valid_places);
 }
 TEST(EfficientNetB0, test_opencl) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
-      Place{TARGET(kARM), PRECISION(kFloat)},
      Place{TARGET(kOpenCL), PRECISION(kFloat)},
+      Place{TARGET(kARM), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kOpenCL), PRECISION(kFloat)}));
+  TestModel(valid_places);
 }
 }  // namespace lite

--- a/lite/api/inceptionv4_test.cc
+++ b/lite/api/inceptionv4_test.cc
@@ -30,14 +30,9 @@ TEST(InceptionV4, test) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_HIGH, FLAGS_threads);
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kARM), PRECISION(kFloat)}});
-                                   Place{TARGET(kARM), PRECISION(kFloat)}});
-  predictor.Build(FLAGS_model_dir,
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
-                  "",
-                  "",
-                  Place{TARGET(kARM), PRECISION(kFloat)},
-                  valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));

--- a/lite/api/lite_api_test_helper.cc
+++ b/lite/api/lite_api_test_helper.cc
@@ -24,24 +24,16 @@ namespace lite {
 const lite::Tensor* RunHvyModel() {
  lite::Predictor predictor;
 #ifndef LITE_WITH_CUDA
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kX86), PRECISION(kFloat)}});
-                                   Place{TARGET(kX86), PRECISION(kFloat)}});
 #else
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW)},
      Place{TARGET(kCUDA), PRECISION(kFloat), DATALAYOUT(kNCHW)},
      Place{TARGET(kCUDA), PRECISION(kAny), DATALAYOUT(kNCHW)},
-      Place{TARGET(kHost), PRECISION(kAny), DATALAYOUT(kNCHW)},
      Place{TARGET(kCUDA), PRECISION(kAny), DATALAYOUT(kAny)},
-      Place{TARGET(kHost), PRECISION(kAny), DATALAYOUT(kAny)},
  });
 #endif
-  predictor.Build(FLAGS_model_dir,
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
-                  "",
-                  "",
-                  Place{TARGET(kX86), PRECISION(kFloat)},  // origin cuda
-                  valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({100, 100})));

--- a/lite/api/mobilenetv1_int8_test.cc
+++ b/lite/api/mobilenetv1_int8_test.cc
@@ -25,13 +25,12 @@
 namespace paddle {
 namespace lite {
-void TestModel(const std::vector<Place>& valid_places,
+void TestModel(const std::vector<Place>& valid_places) {
-               const Place& preferred_place) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_NO_BIND, FLAGS_threads);
  lite::Predictor predictor;
-  predictor.Build(FLAGS_model_dir, "", "", preferred_place, valid_places);
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));
@@ -93,12 +92,11 @@ void TestModel(const std::vector<Place>& valid_places,
 TEST(MobileNetV1, test_arm) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
-      Place{TARGET(kARM), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kInt8)},
+      Place{TARGET(kARM), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kARM), PRECISION(kInt8)}));
+  TestModel(valid_places);
 }
 }  // namespace lite

--- a/lite/api/mobilenetv1_ssd_test.cc
+++ b/lite/api/mobilenetv1_ssd_test.cc
@@ -26,13 +26,12 @@ namespace paddle {
 namespace lite {
 #ifdef LITE_WITH_ARM
-void TestModel(const std::vector<Place>& valid_places,
+void TestModel(const std::vector<Place>& valid_places) {
-               const Place& preferred_place) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_NO_BIND, FLAGS_threads);
  lite::Predictor predictor;
-  predictor.Build(FLAGS_model_dir, "", "", preferred_place, valid_places);
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 300, 300})));
@@ -99,7 +98,6 @@ void TestModel(const std::vector<Place>& valid_places,
 TEST(MobileNetV1_SSD, test_arm) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},
  });

--- a/lite/api/mobilenetv1_test.cc
+++ b/lite/api/mobilenetv1_test.cc
@@ -28,14 +28,13 @@ namespace paddle {
 namespace lite {
 void TestModel(const std::vector<Place>& valid_places,
-               const Place& preferred_place,
               const std::string& model_dir = FLAGS_model_dir,
               bool save_model = false) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_NO_BIND, FLAGS_threads);
  lite::Predictor predictor;
-  predictor.Build(model_dir, "", "", preferred_place, valid_places);
+  predictor.Build(model_dir, "", "", valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));
@@ -103,41 +102,32 @@ void TestModel(const std::vector<Place>& valid_places,
 #ifdef LITE_WITH_NPU
 TEST(MobileNetV1, test_npu) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},
      Place{TARGET(kNPU), PRECISION(kFloat)},
  });
-  TestModel(valid_places,
+  TestModel(valid_places, FLAGS_model_dir, true /* save_model*/);
-            Place({TARGET(kARM), PRECISION(kFloat)}),
-            FLAGS_model_dir,
-            true /* save_model*/);
-  TestModel(valid_places,
+  TestModel(valid_places, FLAGS_optimized_model, false /* save model */);
-            Place({TARGET(kARM), PRECISION(kFloat)}),
-            FLAGS_optimized_model,
-            false /* save model */);
 }
 #endif  // LITE_WITH_NPU
 TEST(MobileNetV1, test_arm) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kARM), PRECISION(kFloat)}));
+  TestModel(valid_places);
 }
 #ifdef LITE_WITH_OPENCL
 TEST(MobileNetV1, test_opencl) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
-      Place{TARGET(kARM), PRECISION(kFloat)},
      Place{TARGET(kOpenCL), PRECISION(kFloat)},
+      Place{TARGET(kARM), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kOpenCL), PRECISION(kFloat)}));
+  TestModel(valid_places);
 }
 #endif  // LITE_WITH_OPENCL

--- a/lite/api/mobilenetv1_yolov3_test.cc
+++ b/lite/api/mobilenetv1_yolov3_test.cc
@@ -26,13 +26,12 @@ namespace paddle {
 namespace lite {
 #ifdef LITE_WITH_ARM
-void TestModel(const std::vector<Place>& valid_places,
+void TestModel(const std::vector<Place>& valid_places) {
-               const Place& preferred_place) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_NO_BIND, FLAGS_threads);
  lite::Predictor predictor;
-  predictor.Build(FLAGS_model_dir, "", "", preferred_place, valid_places);
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 608, 608})));
@@ -106,11 +105,10 @@ void TestModel(const std::vector<Place>& valid_places,
 TEST(MobileNetV1_YoloV3, test_arm) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kARM), PRECISION(kFloat)}));
+  TestModel(valid_places);
 }
 #endif  // LITE_WITH_ARM

--- a/lite/api/mobilenetv2_test.cc
+++ b/lite/api/mobilenetv2_test.cc
@@ -29,14 +29,13 @@ namespace lite {
 #ifdef LITE_WITH_ARM
 void TestModel(const std::vector<Place>& valid_places,
-               const Place& preferred_place,
               const std::string& model_dir = FLAGS_model_dir,
               bool save_model = false) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_NO_BIND, FLAGS_threads);
  lite::Predictor predictor;
-  predictor.Build(model_dir, "", "", preferred_place, valid_places);
+  predictor.Build(model_dir, "", "", valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));
@@ -103,41 +102,32 @@ void TestModel(const std::vector<Place>& valid_places,
 #ifdef LITE_WITH_NPU
 TEST(MobileNetV2, test_npu) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},
      Place{TARGET(kNPU), PRECISION(kFloat)},
  });
-  TestModel(valid_places,
+  TestModel(valid_places, FLAGS_model_dir, true /* save_model*/);
-            Place({TARGET(kARM), PRECISION(kFloat)}),
-            FLAGS_model_dir,
-            true /* save_model*/);
-  TestModel(valid_places,
+  TestModel(valid_places, FLAGS_optimized_model, false /* save model */);
-            Place({TARGET(kARM), PRECISION(kFloat)}),
-            FLAGS_optimized_model,
-            false /* save model */);
 }
 #endif  // LITE_WITH_NPU
 TEST(MobileNetV2, test_arm) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kARM), PRECISION(kFloat)}));
+  TestModel(valid_places);
 }
 #ifdef LITE_WITH_OPENCL
 TEST(MobileNetV2, test_opencl) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
-      Place{TARGET(kARM), PRECISION(kFloat)},
      Place{TARGET(kOpenCL), PRECISION(kFloat)},
+      Place{TARGET(kARM), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kOpenCL), PRECISION(kFloat)}));
+  TestModel(valid_places);
 }
 #endif  // LITE_WITH_OPENCL

--- a/lite/api/model_optimize_tool.cc
+++ b/lite/api/model_optimize_tool.cc
@@ -82,7 +82,6 @@ void Main() {
          target_repr.c_str());
    }
  }
-  valid_places.emplace_back(TARGET(kHost));
  CHECK(!valid_places.empty())
      << "At least one target should be set, should set the "
@@ -90,8 +89,8 @@ void Main() {
  if (FLAGS_prefer_int8_kernel) {
    LOG(WARNING) << "Int8 mode is only support by ARM target";
-    valid_places.push_back(Place{TARGET(kARM), PRECISION(kInt8)});
+    valid_places.insert(valid_places.begin(),
-    config.set_preferred_place(Place{TARGET(kARM), PRECISION(kInt8)});
+                        Place{TARGET(kARM), PRECISION(kInt8)});
  }
  config.set_valid_places(valid_places);

--- a/lite/api/model_run_test_image.cc
+++ b/lite/api/model_run_test_image.cc
@@ -30,16 +30,14 @@ TEST(model, test) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_HIGH, FLAGS_threads);
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kARM), PRECISION(kFloat)},
-                                   Place{TARGET(kARM), PRECISION(kFloat)},
                                   Place{TARGET(kARM), PRECISION(kInt8)}});
  auto precision = PRECISION(kFloat);
  if (FLAGS_int8) {
    precision = PRECISION(kInt8);
  }
-  predictor.Build(
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
-      FLAGS_model_dir, "", "", Place{TARGET(kARM), precision}, valid_places);
  int im_width = FLAGS_im_width;
  int im_height = FLAGS_im_height;
  auto* input_tensor = predictor.GetInput(0);

--- a/lite/api/model_test.cc
+++ b/lite/api/model_test.cc
@@ -36,11 +36,9 @@ void OutputOptModel(const std::string& load_model_dir,
                    const std::vector<std::vector<int64_t>>& input_shapes) {
  lite_api::CxxConfig config;
  config.set_model_dir(load_model_dir);
-  config.set_preferred_place(Place{TARGET(kX86), PRECISION(kFloat)});
  config.set_valid_places({
      Place{TARGET(kX86), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},
-      Place{TARGET(kHost), PRECISION(kFloat)},
  });
  auto predictor = lite_api::CreatePaddlePredictor(config);

--- a/lite/api/ocr_attention_test.cc
+++ b/lite/api/ocr_attention_test.cc
@@ -25,14 +25,12 @@
 namespace paddle {
 namespace lite {
-void TestModel(const std::vector<Place>& valid_places,
+void TestModel(const std::vector<Place>& valid_places, bool use_npu = false) {
-               const Place& preferred_place,
-               bool use_npu = false) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_HIGH, FLAGS_threads);
  lite::Predictor predictor;
-  predictor.Build(FLAGS_model_dir, "", "", preferred_place, valid_places);
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 1, 48, 512})));
@@ -104,11 +102,10 @@ void TestModel(const std::vector<Place>& valid_places,
 TEST(OcrAttention, test_arm) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kARM), PRECISION(kFloat)}));
+  TestModel(valid_places);
 }
 }  // namespace lite

--- a/lite/api/paddle_api.h
+++ b/lite/api/paddle_api.h
@@ -116,14 +116,12 @@ class LITE_API ConfigBase {
 /// CxxConfig is the config for the Full feature predictor.
 class LITE_API CxxConfig : public ConfigBase {
-  Place preferred_place_;
  std::vector<Place> valid_places_;
  std::string model_file_;
  std::string param_file_;
  bool model_from_memory_{false};
 public:
-  void set_preferred_place(const Place& x) { preferred_place_ = x; }
  void set_valid_places(const std::vector<Place>& x) { valid_places_ = x; }
  void set_model_file(const std::string& path) { model_file_ = path; }
  void set_param_file(const std::string& path) { param_file_ = path; }
@@ -136,7 +134,6 @@ class LITE_API CxxConfig : public ConfigBase {
    model_from_memory_ = true;
  }
-  const Place& preferred_place() const { return preferred_place_; }
  const std::vector<Place>& valid_places() const { return valid_places_; }
  std::string model_file() const { return model_file_; }
  std::string param_file() const { return param_file_; }

--- a/lite/api/paddle_api_test.cc
+++ b/lite/api/paddle_api_test.cc
@@ -28,7 +28,6 @@ namespace lite_api {
 TEST(CxxApi, run) {
  lite_api::CxxConfig config;
  config.set_model_dir(FLAGS_model_dir);
-  config.set_preferred_place(Place{TARGET(kX86), PRECISION(kFloat)});
  config.set_valid_places({
      Place{TARGET(kX86), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},

--- a/lite/api/resnet18_test.cc
+++ b/lite/api/resnet18_test.cc
@@ -28,14 +28,9 @@ namespace lite {
 #ifdef LITE_WITH_ARM
 TEST(ResNet18, test) {
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kARM), PRECISION(kFloat)}});
-                                   Place{TARGET(kARM), PRECISION(kFloat)}});
-  predictor.Build(FLAGS_model_dir,
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
-                  "",
-                  "",
-                  Place{TARGET(kARM), PRECISION(kFloat)},
-                  valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));

--- a/lite/api/resnet50_test.cc
+++ b/lite/api/resnet50_test.cc
@@ -26,13 +26,12 @@ namespace paddle {
 namespace lite {
 #ifdef LITE_WITH_ARM
-void TestModel(const std::vector<Place>& valid_places,
+void TestModel(const std::vector<Place>& valid_places) {
-               const Place& preferred_place) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_HIGH, FLAGS_threads);
  lite::Predictor predictor;
-  predictor.Build(FLAGS_model_dir, "", "", preferred_place, valid_places);
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));
@@ -82,22 +81,20 @@ void TestModel(const std::vector<Place>& valid_places,
 TEST(ResNet50, test_arm) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kARM), PRECISION(kFloat)}));
+  TestModel(valid_places);
 }
 #ifdef LITE_WITH_OPENCL
 TEST(ResNet50, test_opencl) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
-      Place{TARGET(kARM), PRECISION(kFloat)},
      Place{TARGET(kOpenCL), PRECISION(kFloat)},
+      Place{TARGET(kARM), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kOpenCL), PRECISION(kFloat)}));
+  TestModel(valid_places);
 }
 #endif  // LITE_WITH_OPENCL

--- a/lite/api/resnet50_test_fpga.cc
+++ b/lite/api/resnet50_test_fpga.cc
@@ -29,8 +29,7 @@ namespace lite {
 TEST(ResNet50, test) {
  lite::Predictor predictor;
  std::vector<Place> valid_places(
-      {Place{TARGET(kFPGA), PRECISION(kFP16), DATALAYOUT(kNHWC)},
+      {Place{TARGET(kFPGA), PRECISION(kFP16), DATALAYOUT(kNHWC)}});
-       Place{TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNHWC)}});
  predictor.Build(FLAGS_model_dir,
                  "",

--- a/lite/api/shufflenetv2_test.cc
+++ b/lite/api/shufflenetv2_test.cc
@@ -25,13 +25,12 @@
 namespace paddle {
 namespace lite {
-void TestModel(const std::vector<Place>& valid_places,
+void TestModel(const std::vector<Place>& valid_places) {
-               const Place& preferred_place) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_HIGH, FLAGS_threads);
  lite::Predictor predictor;
-  predictor.Build(FLAGS_model_dir, "", "", preferred_place, valid_places);
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim((std::vector<DDim::value_type>({1, 3, 224, 224}))));
@@ -80,12 +79,11 @@ void TestModel(const std::vector<Place>& valid_places,
 TEST(ShuffleNetV2, test_arm) {
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
      Place{TARGET(kARM), PRECISION(kFloat)},
      // Place{TARGET(kOpenCL), PRECISION(kFloat)},
  });
-  TestModel(valid_places, Place({TARGET(kARM), PRECISION(kFloat)}));
+  TestModel(valid_places);
 }
 }  // namespace lite

--- a/lite/api/test_googlenet_lite.cc
+++ b/lite/api/test_googlenet_lite.cc
@@ -45,13 +45,11 @@ namespace lite {
 #ifdef LITE_WITH_X86
 TEST(CXXApi, test_lite_googlenet) {
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kX86), PRECISION(kFloat)}});
-                                   Place{TARGET(kX86), PRECISION(kFloat)}});
  //  LOG(INFO)<<"FLAGS_eval_googlenet_dir:"<<FLAGS_test_lite_googlenet_dir;
  std::string model_dir = FLAGS_model_dir;
-  predictor.Build(
+  predictor.Build(model_dir, "", "", valid_places);
-      model_dir, "", "", Place{TARGET(kX86), PRECISION(kFloat)}, valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));

--- a/lite/api/test_inceptionv4_lite_x86.cc
+++ b/lite/api/test_inceptionv4_lite_x86.cc
@@ -43,8 +43,7 @@ namespace lite {
 TEST(InceptionV4, test_inceptionv4_lite_x86) {
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kX86), PRECISION(kFloat)}});
-                                   Place{TARGET(kX86), PRECISION(kFloat)}});
  //  LOG(INFO)<<"FLAGS_eval_googlenet_dir:"<<FLAGS_test_lite_googlenet_dir;
  std::string model_dir = FLAGS_model_dir;
@@ -55,12 +54,7 @@ TEST(InceptionV4, test_inceptionv4_lite_x86) {
                                   "io_copy_kernel_pick_pass",
                                   "variable_place_inference_pass",
                                   "runtime_context_assign_pass"});
-  predictor.Build(model_dir,
+  predictor.Build(model_dir, "", "", valid_places, passes);
-                  "",
-                  "",
-                  Place{TARGET(kX86), PRECISION(kFloat)},
-                  valid_places,
-                  passes);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));

--- a/lite/api/test_mobilenetv1_lite_x86.cc
+++ b/lite/api/test_mobilenetv1_lite_x86.cc
@@ -43,8 +43,7 @@ namespace lite {
 TEST(Mobilenet_v1, test_mobilenetv1_lite_x86) {
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kX86), PRECISION(kFloat)}});
-                                   Place{TARGET(kX86), PRECISION(kFloat)}});
  std::string model_dir = FLAGS_model_dir;
  std::vector<std::string> passes({"static_kernel_pick_pass",
@@ -54,12 +53,7 @@ TEST(Mobilenet_v1, test_mobilenetv1_lite_x86) {
                                   "io_copy_kernel_pick_pass",
                                   "variable_place_inference_pass",
                                   "runtime_context_assign_pass"});
-  predictor.Build(model_dir,
+  predictor.Build(model_dir, "", "", valid_places, passes);
-                  "",
-                  "",
-                  Place{TARGET(kX86), PRECISION(kFloat)},
-                  valid_places,
-                  passes);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));
  auto* data = input_tensor->mutable_data<float>();

--- a/lite/api/test_mobilenetv2_lite_x86.cc
+++ b/lite/api/test_mobilenetv2_lite_x86.cc
@@ -44,8 +44,8 @@ namespace lite {
 TEST(Mobilenet_v2, test_mobilenetv2_lite_x86) {
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kX86), PRECISION(kFloat)},
-                                   Place{TARGET(kX86), PRECISION(kFloat)}});
+                                   Place{TARGET(kHost), PRECISION(kFloat)}});
  //  LOG(INFO)<<"FLAGS_eval_googlenet_dir:"<<FLAGS_test_lite_googlenet_dir;
  std::string model_dir = FLAGS_model_dir;
@@ -56,12 +56,7 @@ TEST(Mobilenet_v2, test_mobilenetv2_lite_x86) {
                                   "io_copy_kernel_pick_pass",
                                   "variable_place_inference_pass",
                                   "runtime_context_assign_pass"});
-  predictor.Build(model_dir,
+  predictor.Build(model_dir, "", "", valid_places, passes);
-                  "",
-                  "",
-                  Place{TARGET(kX86), PRECISION(kFloat)},
-                  valid_places,
-                  passes);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 224, 224})));

--- a/lite/api/unet_test.cc
+++ b/lite/api/unet_test.cc
@@ -30,14 +30,9 @@ TEST(unet, test) {
  DeviceInfo::Init();
  DeviceInfo::Global().SetRunMode(lite_api::LITE_POWER_HIGH, FLAGS_threads);
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kARM), PRECISION(kFloat)}});
-                                   Place{TARGET(kARM), PRECISION(kFloat)}});
-  predictor.Build(FLAGS_model_dir,
+  predictor.Build(FLAGS_model_dir, "", "", valid_places);
-                  "",
-                  "",
-                  Place{TARGET(kARM), PRECISION(kFloat)},
-                  valid_places);
  auto* input_tensor = predictor.GetInput(0);
  input_tensor->Resize(DDim(std::vector<DDim::value_type>({1, 3, 512, 512})));

--- a/lite/core/mir/fusion/fc_fuse_pass_test.cc
+++ b/lite/core/mir/fusion/fc_fuse_pass_test.cc
@@ -30,16 +30,12 @@ namespace mir {
 TEST(fc_fuse_pass, fuse_test) {
  lite::Predictor predictor;
 #ifndef LITE_WITH_CUDA
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kX86), PRECISION(kFloat)}});
-                                   Place{TARGET(kX86), PRECISION(kFloat)}});
 #else
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat), DATALAYOUT(kNCHW)},
      Place{TARGET(kCUDA), PRECISION(kFloat), DATALAYOUT(kNCHW)},
      Place{TARGET(kCUDA), PRECISION(kAny), DATALAYOUT(kNCHW)},
-      Place{TARGET(kHost), PRECISION(kAny), DATALAYOUT(kNCHW)},
      Place{TARGET(kCUDA), PRECISION(kAny), DATALAYOUT(kAny)},
-      Place{TARGET(kHost), PRECISION(kAny), DATALAYOUT(kAny)},
  });
 #endif
@@ -72,8 +68,7 @@ TEST(fc_fuse_pass, fuse_test) {
 #ifndef LITE_WITH_LIGHT_WEIGHT_FRAMEWORK
 TEST(fc_fuse_pass, save_model_test) {
  lite::Predictor predictor;
-  std::vector<Place> valid_places({Place{TARGET(kHost), PRECISION(kFloat)},
+  std::vector<Place> valid_places({Place{TARGET(kX86), PRECISION(kFloat)}});
-                                   Place{TARGET(kX86), PRECISION(kFloat)}});
  predictor.Build(FLAGS_model_dir,
                  "",
                  "",

--- a/lite/core/mir/static_kernel_pick_pass.cc
+++ b/lite/core/mir/static_kernel_pick_pass.cc
@@ -24,8 +24,8 @@ namespace paddle {
 namespace lite {
 namespace mir {
-bool KernelScoreCmp(const std::pair<size_t, std::unique_ptr<KernelBase>>& a,
+bool KernelScoreCmp(const std::pair<float, std::unique_ptr<KernelBase>>& a,
-                    const std::pair<size_t, std::unique_ptr<KernelBase>>& b) {
+                    const std::pair<float, std::unique_ptr<KernelBase>>& b) {
  return a.first > b.first;
 }
@@ -44,12 +44,12 @@ void StaticKernelPickPass::Apply(const std::unique_ptr<SSAGraph>& graph) {
    auto& instruct = node.AsStmt();
    // Get candidate kernels
-    std::vector<std::pair<size_t, std::unique_ptr<KernelBase>>> scored;
+    std::vector<std::pair<float, std::unique_ptr<KernelBase>>> scored;
    CHECK(!instruct.kernels().empty()) << "No kernels found for "
                                       << instruct.op_type();
    VLOG(4) << "instruct.kernels().size():" << instruct.kernels().size();
    for (auto&& kernel : instruct.kernels()) {
-      size_t score = KernelGrade(*kernel);
+      float score = KernelGrade(*kernel, graph->valid_places());
      VLOG(4) << "kernel->summary():" << kernel->summary()
              << " score:" << score;
      scored.emplace_back(score, std::move(kernel));
@@ -99,7 +99,7 @@ void StaticKernelPickPass::Apply(const std::unique_ptr<SSAGraph>& graph) {
        instruct.ResetOp(update_desc, graph->valid_places());
        scored.clear();
        for (auto&& kernel : instruct.kernels()) {
-          size_t score = KernelGrade(*kernel);
+          float score = KernelGrade(*kernel, graph->valid_places());
          scored.emplace_back(score, std::move(kernel));
        }
        std::sort(scored.begin(), scored.end(), KernelScoreCmp);

--- a/lite/core/mir/static_kernel_pick_pass.h
+++ b/lite/core/mir/static_kernel_pick_pass.h
@@ -16,6 +16,7 @@
 #include <limits>
 #include <memory>
+#include <vector>
 #include "lite/core/mir/pass.h"
 #include "lite/core/types.h"
@@ -38,8 +39,6 @@ class StaticKernelPickPass : public mir::StmtPass {
 public:
  void Apply(const std::unique_ptr<SSAGraph>& graph) override;
-  void SetPreferPlace(const Place& place) { place_ = place; }
-  const Place& place() const { return place_; }
  const core::KernelPickFactor& kernel_pick_factors() const {
    return kernel_pick_factors_;
  }
@@ -49,39 +48,63 @@ class StaticKernelPickPass : public mir::StmtPass {
 private:
  // Score the kernel.
-  size_t KernelGrade(const lite::KernelBase& kernel) {
+  size_t KernelGrade(const lite::KernelBase& kernel,
-    size_t score{};
+                     const std::vector<Place>& places) {
+    CHECK_GT(places.size(), 0) << "valid_places is empty.";
+    float final_score{-1.};
+    Place winner_place{places[0]};
    const int kMax =
        std::numeric_limits<core::KernelPickFactor::value_type>::max();
-    VLOG(4) << "[score s1]:" << score;
+    size_t place_size = places.size();
-    // The more important factor comes first
-    if (kernel_pick_factors_.IsTargetConsidered() &&
+    // NOTE: We compare kernel's place with place in valid_places to select the
-        (place().target == kernel.target() || kernel.target() == TARGET(kAny) ||
+    // best match place
-         place().target == TARGET(kAny))) {
+    //       The place's order in valid_places array decide the user's
-      score +=
+    //       preference
-          kMax / static_cast<int>(core::KernelPickFactor::Factor::TargetFirst);
+    // final_score = weight * socre
-    }
+    // weight: The weight is compute with (valid_places.size() - i) /
-    VLOG(4) << "[score s2]:" << score;
+    // valid_places.size() as default.
-    if (kernel_pick_factors_.IsPrecisionConsidered() &&
+    //         where i is the place's index in valid_places array.
-        (place().precision == kernel.precision() ||
+    // score:  score is the weighted sum of target、percision and layout
-         kernel.precision() == PRECISION(kAny) ||
+    for (int i = 0; i < place_size; ++i) {
-         place().precision == PRECISION(kAny))) {
+      const auto& place = places[i];
-      score += kMax /
+      float weight = static_cast<float>(place_size - i) / place_size;
-               static_cast<int>(core::KernelPickFactor::Factor::PrecisionFirst);
+      size_t score{};
-    }
+      // The more important factor comes first
-    VLOG(4) << "[score s3]:" << score;
+      if (kernel_pick_factors_.IsTargetConsidered() &&
-    if (kernel_pick_factors_.IsDataLayoutConsidered() &&
+          (place.target == kernel.target() || kernel.target() == TARGET(kAny) ||
-        (place().layout == kernel.layout() ||
+           place.target == TARGET(kAny))) {
-         kernel.layout() == DATALAYOUT(kAny) ||
+        score += kMax /
-         place().layout == DATALAYOUT(kAny))) {
+                 static_cast<int>(core::KernelPickFactor::Factor::TargetFirst);
-      score += kMax / static_cast<int>(
+      }
-                          core::KernelPickFactor::Factor::DataLayoutFirst);
+      VLOG(4) << "[score s1]:" << score;
+      if (kernel_pick_factors_.IsPrecisionConsidered() &&
+          (place.precision == kernel.precision() ||
+           kernel.precision() == PRECISION(kAny) ||
+           place.precision == PRECISION(kAny))) {
+        score += kMax / static_cast<int>(
+                            core::KernelPickFactor::Factor::PrecisionFirst);
+      }
+      VLOG(4) << "[score s2]:" << score;
+      if (kernel_pick_factors_.IsDataLayoutConsidered() &&
+          (place.layout == kernel.layout() ||
+           kernel.layout() == DATALAYOUT(kAny) ||
+           place.layout == DATALAYOUT(kAny))) {
+        score += kMax / static_cast<int>(
+                            core::KernelPickFactor::Factor::DataLayoutFirst);
+      }
+      VLOG(4) << "[score s3]:" << score;
+      if (weight * score > final_score) {
+        final_score = weight * score;
+        winner_place = place;
+      }
    }
-    VLOG(4) << "[score s4(final)]:" << score;
+    VLOG(4) << "[score(final)]:" << final_score;
    VLOG(4) << "-------- pick summary --------";
-    VLOG(4) << " ===> place():" << PrecisionToStr(place().precision) << " "
+    VLOG(4) << " ===> place():" << PrecisionToStr(winner_place.precision) << " "
-            << DataLayoutToStr(place().layout) << " "
+            << DataLayoutToStr(winner_place.layout) << " "
-            << TargetToStr(place().target);
+            << TargetToStr(winner_place.target);
    VLOG(4) << " ===> kernel.place():"
            << PrecisionToStr(kernel.place().precision) << " "
            << DataLayoutToStr(kernel.place().layout) << " "
@@ -89,20 +112,18 @@ class StaticKernelPickPass : public mir::StmtPass {
    VLOG(4) << "kernel.op_type():" << kernel.op_type();
    VLOG(4) << "picker tactic " << kernel_pick_factors_;
    VLOG(4) << "kernel place " << kernel.place().DebugString();
-    VLOG(4) << "picker place " << place().DebugString();
+    VLOG(4) << "picker place " << winner_place.DebugString();
-    VLOG(4) << "score " << score;
    VLOG(4) << "------------------------------";
    // The data layout is not considered, for the input and output arguments
    // might have different data layout.
    // TODO(Superjomn) reconsider the idea of taking the data layout as a kernel
    // specification.
-    return score;
+    return final_score;
  }
 private:
  core::KernelPickFactor kernel_pick_factors_;
-  Place place_;
 };
 }  // namespace mir

--- a/lite/core/mir/subgraph/generate_npu_program_pass_test.cc
+++ b/lite/core/mir/subgraph/generate_npu_program_pass_test.cc
@@ -106,7 +106,6 @@ std::shared_ptr<lite_api::PaddlePredictor> TestModel(
    const std::string& model_dir,
    const std::string& model_file,
    const std::string& params_file,
-    const lite_api::Place& preferred_place,
    const std::vector<lite_api::Place>& valid_places,
    const std::vector<std::vector<int64_t>>& input_tensor_shape,
    const std::string& optimized_model_dir) {
@@ -115,7 +114,6 @@ std::shared_ptr<lite_api::PaddlePredictor> TestModel(
  cxx_config.set_model_dir(model_dir);
  cxx_config.set_model_file(model_file);
  cxx_config.set_param_file(params_file);
-  cxx_config.set_preferred_place(preferred_place);
  cxx_config.set_valid_places(valid_places);
  auto predictor = lite_api::CreatePaddlePredictor(cxx_config);
  FillInputTensor(predictor, input_tensor_shape, 1);
@@ -151,9 +149,7 @@ TEST(NPUSubgraph, compare) {
      TestModel(FLAGS_model_dir,
                FLAGS_model_file,
                FLAGS_params_file,
-                lite_api::Place{TARGET(kARM), PRECISION(kFloat)},
+                {lite_api::Place{TARGET(kARM), PRECISION(kFloat)}},
-                {lite_api::Place{TARGET(kHost), PRECISION(kFloat)},
-                 lite_api::Place{TARGET(kARM), PRECISION(kFloat)}},
                input_tensor_shape,
                FLAGS_optimized_model_dir + "/CPU");
  // generate and run optimized NPU model
@@ -162,9 +158,7 @@ TEST(NPUSubgraph, compare) {
      TestModel(FLAGS_model_dir,
                FLAGS_model_file,
                FLAGS_params_file,
-                lite_api::Place{TARGET(kARM), PRECISION(kFloat)},
+                {lite_api::Place{TARGET(kARM), PRECISION(kFloat)},
-                {lite_api::Place{TARGET(kHost), PRECISION(kFloat)},
-                 lite_api::Place{TARGET(kARM), PRECISION(kFloat)},
                 lite_api::Place{TARGET(kNPU), PRECISION(kFloat)}},
                input_tensor_shape,
                FLAGS_optimized_model_dir + "/NPU");

--- a/lite/core/mir/variable_place_inference_pass_test.cc
+++ b/lite/core/mir/variable_place_inference_pass_test.cc
@@ -63,18 +63,6 @@ TEST(variable_place_inference_pass, test) {
      "type_target_cast_pass",          //
  });
-  Place prefered_place{
-#ifdef PADDLE_WITH_CUDA
-      TARGET(kCUDA), PRECISION(kFloat), DATALAYOUT(kNCHW),
-#else
-#ifdef PADDLE_WITH_ARM
-      TARGET(kARM), PRECISION(kFloat), DATALAYOUT(kNCHW),
-#else   // X86
-      TARGET(kX86), PRECISION(kFloat), DATALAYOUT(kNCHW),
-#endif  // ARM
-#endif
-  };
-  optimizer.KernelPickPreferPlace(prefered_place);
  optimizer.Run(std::move(program), places, factor, passes);
 }

--- a/lite/core/optimizer.h
+++ b/lite/core/optimizer.h
@@ -117,13 +117,6 @@ class Optimizer {
    exec_scope_ = program.exec_scope();
  }
-  void KernelPickPreferPlace(const Place& place) {
-    auto* pass = mir::PassManager::Global().LookUp<mir::StaticKernelPickPass>(
-        "static_kernel_pick_pass");
-    CHECK(pass);
-    pass->SetPreferPlace(place);
-  }
  const lite::Scope* exec_scope() const { return exec_scope_; }
  // Generate a new program based on the mir graph.

--- a/lite/demo/cxx/mobile_full/mobilenetv1_full_api.cc
+++ b/lite/demo/cxx/mobile_full/mobilenetv1_full_api.cc
@@ -38,10 +38,8 @@ void RunModel() {
  config.set_model_dir(FLAGS_model_dir);
  std::vector<Place> valid_places{Place{TARGET(kARM), PRECISION(kFloat)}};
  if (FLAGS_prefer_int8_kernel) {
-    valid_places.push_back(Place{TARGET(kARM), PRECISION(kInt8)});
+    valid_places.insert(valid_places.begin(),
-    config.set_preferred_place(Place{TARGET(kARM), PRECISION(kInt8)});
+                        Place{TARGET(kARM), PRECISION(kInt8)});
-  } else {
-    config.set_preferred_place(Place{TARGET(kARM), PRECISION(kFloat)});
  }
  config.set_valid_places(valid_places);

--- a/lite/gen_code/gen_code.h
+++ b/lite/gen_code/gen_code.h
@@ -102,7 +102,7 @@ class Module {
  void AddValidPlaceDecl() {
    // clang-format off
-    Line("std::vector<lite::Place> valid_places({lite::Place({TARGET(kX86), PRECISION(kFloat), DATALAYOUT(kNCHW)}), lite::Place({TARGET(kHost), PRECISION(kAny), DATALAYOUT(kAny)})});");  // NOLINT
+    Line("std::vector<lite::Place> valid_places({lite::Place({TARGET(kX86), PRECISION(kFloat), DATALAYOUT(kNCHW)})});");  // NOLINT
    // clang-format on
  }

--- a/lite/tools/debug/model_debug_tool.cc
+++ b/lite/tools/debug/model_debug_tool.cc
@@ -35,7 +35,6 @@ void Run(DebugConfig* conf) {
 #endif
  lite::Predictor predictor;
  std::vector<Place> valid_places({
-      Place{TARGET(kHost), PRECISION(kFloat)},
 #ifdef LITE_WITH_ARM
      Place{TARGET(kARM), PRECISION(kFloat)},
 #endif
@@ -60,23 +59,7 @@ void Run(DebugConfig* conf) {
      "runtime_context_assign_pass",
  }};
-  predictor.Build(conf->model_dir,
+  predictor.Build(conf->model_dir, "", "", valid_places, passes);
-                  "",
-                  "",
-#ifdef LITE_WITH_ARM
-                  Place{TARGET(kARM), PRECISION(kFloat)},
-#endif
-#ifdef LITE_WITH_X86
-                  Place{TARGET(kX86), PRECISION(kFloat)},
-#endif
-#ifdef LITE_WITH_FPGA
-                  Place{TARGET(kFPGA), PRECISION(kFloat)},
-#endif
-#ifdef LITE_WITH_CUDA
-                  Place{TARGET(kCUDA), PRECISION(kFloat)},
-#endif
-                  valid_places,
-                  passes);
  predictor.GenRuntimeProgram();
  auto& instructions = predictor.runtime_program().instructions();