Resolve operator and allocator registering static variable issue

mace/core/BUILD

@@ -10,38 +10,27 @@ licenses(["notice"])  # Apache 2.0
 load("//mace:mace.bzl", "if_android", "if_profiling_enabled", "if_embed_binary_program")
 
 cc_library(
-    name = "opencl_runtime",
+    name = "core",
     srcs = glob([
+        "*.cc",
         "runtime/opencl/*.cc",
     ]),
     hdrs = glob([
+        "*.h",
+        "public/*.h",
         "runtime/opencl/cl2.hpp",
         "runtime/opencl/*.h",
     ]),
     copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"] +
             if_profiling_enabled(["-DMACE_OPENCL_PROFILING"]) +
             if_embed_binary_program(["-DMACE_EMBED_BINARY_PROGRAM"]),
-    linkopts = ["-ldl"],
+    linkopts = if_android(["-pie", "-ldl"]),
     deps = [
-        ":core",
+        "//mace/utils:utils_hdrs",
         "//mace/utils:logging",
         "//mace/utils:tuner",
         "@opencl_headers//:opencl20_headers",
     ],
-    alwayslink = 1,
-)
-
-
-cc_library(
-    name = "core",
-    srcs = glob(["*.cc"]),
-    hdrs = glob(["*.h", "public/*.h"]),
-    copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
-    linkopts = if_android(["-pie"]),
-    deps = [
-        "//mace/utils:utils_hdrs",
-        "//mace/utils:logging",
-    ],
 )
 
 cc_library(

mace/core/allocator.cc

@@ -3,6 +3,7 @@
 //
 
 #include "mace/core/allocator.h"
+#include "mace/core/runtime/opencl/opencl_allocator.h"
 
 namespace mace {
 
@@ -22,5 +23,6 @@ Allocator *GetDeviceAllocator(DeviceType type) {
 
 MACE_REGISTER_ALLOCATOR(DeviceType::CPU, new CPUAllocator());
 MACE_REGISTER_ALLOCATOR(DeviceType::NEON, new CPUAllocator());
+MACE_REGISTER_ALLOCATOR(DeviceType::OPENCL, new OpenCLAllocator());
 
 }  // namespace mace

mace/core/mace.cc

@@ -5,6 +5,7 @@
 #include "mace/core/public/mace.h"
 #include "mace/core/types.h"
 #include "mace/core/net.h"
+#include "mace/core/operator.h"
 #include "mace/core/workspace.h"
 #include "mace/utils/logging.h"
 
@@ -481,17 +482,19 @@ const OperatorDef &NetDef::op(const int idx) const {
 
 // Mace Engine
 MaceEngine::MaceEngine(const NetDef *net_def, DeviceType device_type):
-    device_type_(device_type), ws_(new Workspace()), net_(nullptr) {
+    op_registry_(new OperatorRegistry()), device_type_(device_type),
+    ws_(new Workspace()), net_(nullptr) {
 
   ws_->LoadModelTensor(*net_def, device_type);
 
   // Init model
-  auto net = CreateNet(*net_def, ws_.get(), device_type, NetMode::INIT);
+  auto net = CreateNet(op_registry_, *net_def, ws_.get(),
+                       device_type, NetMode::INIT);
   if(!net->Run()) {
     LOG(FATAL) << "Net init run failed";
   }
   ws_->CreateTensor("mace_input_node:0", GetDeviceAllocator(device_type_), DT_FLOAT);
-  net_ = std::move(CreateNet(*net_def, ws_.get(), device_type));
+  net_ = std::move(CreateNet(op_registry_, *net_def, ws_.get(), device_type));
 }
 MaceEngine::~MaceEngine() = default;
 bool MaceEngine::Run(const float *input,

mace/core/net.cc

@@ -3,22 +3,24 @@
 //
 
 #include "mace/core/net.h"
-#include "mace/core/operator.h"
 #include "mace/core/workspace.h"
 #include "mace/utils/utils.h"
 
 namespace mace {
 
-NetBase::NetBase(const std::shared_ptr<const NetDef> &net_def,
+NetBase::NetBase(const std::shared_ptr<const OperatorRegistry> op_registry,
+                 const std::shared_ptr<const NetDef> net_def,
                  Workspace *ws,
                  DeviceType type)
-    : name_(net_def->name()) {}
+    : op_registry_(op_registry), name_(net_def->name()) {}
 
-SimpleNet::SimpleNet(const std::shared_ptr<const NetDef> &net_def,
+SimpleNet::SimpleNet(const std::shared_ptr<const OperatorRegistry> op_registry,
+                     const std::shared_ptr<const NetDef> net_def,
                      Workspace *ws,
                      DeviceType type,
                      const NetMode mode)
-    : NetBase(net_def, ws, type), device_type_(type){
+    :  NetBase(op_registry, net_def, ws, type),
+      device_type_(type) {
   VLOG(1) << "Constructing SimpleNet " << net_def->name();
   for (int idx = 0; idx < net_def->op_size(); ++idx) {
     const auto &operator_def = net_def->op(idx);
@@ -26,7 +28,7 @@ SimpleNet::SimpleNet(const std::shared_ptr<const NetDef> &net_def,
             << operator_def.type();
     std::unique_ptr<OperatorBase> op{nullptr};
     OperatorDef temp_def(operator_def);
-    op = CreateOperator(temp_def, ws, type, mode);
+    op = op_registry->CreateOperator(temp_def, ws, type, mode);
     if (op) {
       operators_.emplace_back(std::move(op));
     }
@@ -62,9 +64,8 @@ bool SimpleNet::Run(RunMetadata *run_metadata) {
     }
 
     if (run_metadata != nullptr) {
-      OperatorStats op_stats = { op->debug_def().name(),
-                                 op->debug_def().type(),
-                                 call_stats };
+      OperatorStats op_stats = {op->debug_def().name(), op->debug_def().type(),
+                                call_stats};
       run_metadata->op_stats.emplace_back(op_stats);
     }
 
@@ -80,19 +81,23 @@ bool SimpleNet::Run(RunMetadata *run_metadata) {
   return true;
 }
 
-unique_ptr<NetBase> CreateNet(const NetDef &net_def,
+std::unique_ptr<NetBase> CreateNet(
+    const std::shared_ptr<const OperatorRegistry> op_registry,
+    const NetDef &net_def,
     Workspace *ws,
     DeviceType type,
     const NetMode mode) {
   std::shared_ptr<NetDef> tmp_net_def(new NetDef(net_def));
-  return CreateNet(tmp_net_def, ws, type, mode);
+  return CreateNet(op_registry, tmp_net_def, ws, type, mode);
 }
 
-unique_ptr<NetBase> CreateNet(const std::shared_ptr<const NetDef> &net_def,
+std::unique_ptr<NetBase> CreateNet(
+    const std::shared_ptr<const OperatorRegistry> op_registry,
+    const std::shared_ptr<const NetDef> net_def,
     Workspace *ws,
     DeviceType type,
     const NetMode mode) {
-  unique_ptr<NetBase> net(new SimpleNet(net_def, ws, type, mode));
+  unique_ptr<NetBase> net(new SimpleNet(op_registry, net_def, ws, type, mode));
   return net;
 }
 

mace/core/net.h

@@ -6,6 +6,7 @@
 #define MACE_CORE_NET_H_
 
 #include "mace/core/common.h"
+#include "mace/core/operator.h"
 #include "mace/core/public/mace.h"
 
 namespace mace {
@@ -16,7 +17,8 @@ class Workspace;
 
 class NetBase {
  public:
-  NetBase(const std::shared_ptr<const NetDef> &net_def,
+  NetBase(const std::shared_ptr<const OperatorRegistry> op_registry,
+          const std::shared_ptr<const NetDef> net_def,
           Workspace *ws,
           DeviceType type);
   virtual ~NetBase() noexcept {}
@@ -27,13 +29,15 @@ class NetBase {
 
  protected:
   string name_;
+  const std::shared_ptr<const OperatorRegistry> op_registry_;
 
   DISABLE_COPY_AND_ASSIGN(NetBase);
 };
 
 class SimpleNet : public NetBase {
  public:
-  SimpleNet(const std::shared_ptr<const NetDef> &net_def,
+  SimpleNet(const std::shared_ptr<const OperatorRegistry> op_registry,
+            const std::shared_ptr<const NetDef> net_def,
             Workspace *ws,
             DeviceType type,
             const NetMode mode = NetMode::NORMAL);
@@ -47,11 +51,15 @@ class SimpleNet : public NetBase {
   DISABLE_COPY_AND_ASSIGN(SimpleNet);
 };
 
-unique_ptr<NetBase> CreateNet(const NetDef &net_def,
+std::unique_ptr<NetBase> CreateNet(
+    const std::shared_ptr<const OperatorRegistry> op_registry,
+    const NetDef &net_def,
     Workspace *ws,
     DeviceType type,
     const NetMode mode = NetMode::NORMAL);
-unique_ptr<NetBase> CreateNet(const std::shared_ptr<const NetDef> &net_def,
+std::unique_ptr<NetBase> CreateNet(
+    const std::shared_ptr<const OperatorRegistry> op_registry,
+    const std::shared_ptr<const NetDef> net_def,
     Workspace *ws,
     DeviceType type,
     const NetMode mode = NetMode::NORMAL);

mace/core/operator.cc

@@ -2,12 +2,19 @@
 // Copyright (c) 2017 XiaoMi All rights reserved.
 //
 
+#include <sstream>
+
 #include "mace/core/operator.h"
 
 namespace mace {
 
+OperatorBase::OperatorBase(const OperatorDef &operator_def, Workspace *ws)
+    : operator_ws_(ws),
+      operator_def_(std::make_shared<OperatorDef>(operator_def)) {}
 
-OpKeyBuilder::OpKeyBuilder(const char *op_name): op_name_(op_name) {}
+OpKeyBuilder::OpKeyBuilder(const char *op_name) : op_name_(op_name) {}
+
+OpKeyBuilder &OpKeyBuilder::Device(DeviceType device) { device_type_ = device; }
 
 OpKeyBuilder &OpKeyBuilder::TypeConstraint(const char *attr_name,
                                            const DataType allowed) {
@@ -17,61 +24,72 @@ OpKeyBuilder &OpKeyBuilder::TypeConstraint(const char *attr_name,
 
 const std::string OpKeyBuilder::Build() {
   static const std::vector<std::string> type_order = {"T"};
-  std::string key = op_name_;
+  std::stringstream ss;
+  ss << op_name_;
+  ss << device_type_;
   for (auto type : type_order) {
-    key += type + "_" + DataTypeToString(type_constraint_[type]);
+    ss << type << "_" << DataTypeToString(type_constraint_[type]);
   }
-  return key;
-}
 
-std::map<int32_t, OperatorRegistry *> *gDeviceTypeRegistry() {
-  static std::map<int32_t, OperatorRegistry *> g_device_type_registry;
-  return &g_device_type_registry;
+  return ss.str();
 }
 
-MACE_DEFINE_REGISTRY(CPUOperatorRegistry,
-                     OperatorBase,
-                     const OperatorDef &,
-                     Workspace *);
-MACE_REGISTER_DEVICE_TYPE(DeviceType::CPU, CPUOperatorRegistry);
-
-MACE_DEFINE_REGISTRY(NEONOperatorRegistry,
-                     OperatorBase,
-                     const OperatorDef &,
-                     Workspace *);
-MACE_REGISTER_DEVICE_TYPE(DeviceType::NEON, NEONOperatorRegistry);
-
-MACE_DEFINE_REGISTRY(OPENCLOperatorRegistry,
-                     OperatorBase,
-                     const OperatorDef &,
-                     Workspace *);
-MACE_REGISTER_DEVICE_TYPE(DeviceType::OPENCL, OPENCLOperatorRegistry);
-
-unique_ptr<OperatorBase> CreateOperator(const OperatorDef &operator_def,
+std::unique_ptr<OperatorBase> OperatorRegistry::CreateOperator(
+    const OperatorDef &operator_def,
     Workspace *ws,
     DeviceType type,
-                                        const NetMode mode) {
-  OperatorRegistry *registry = gDeviceTypeRegistry()->at(type);
-  const int dtype = ArgumentHelper::GetSingleArgument<OperatorDef, int>(operator_def,
-                                                                        "T",
-                                                                        static_cast<int>(DT_FLOAT));
-  const int op_mode_i= ArgumentHelper::GetSingleArgument<OperatorDef, int>(operator_def,
-                                                                        "mode",
-                                                                        static_cast<int>(NetMode::NORMAL));
+    const NetMode mode) const {
+  const int dtype = ArgumentHelper::GetSingleArgument<OperatorDef, int>(
+      operator_def, "T", static_cast<int>(DT_FLOAT));
+  const int op_mode_i = ArgumentHelper::GetSingleArgument<OperatorDef, int>(
+      operator_def, "mode", static_cast<int>(NetMode::NORMAL));
   const NetMode op_mode = static_cast<NetMode>(op_mode_i);
   if (op_mode == mode) {
-    return registry->Create(OpKeyBuilder(operator_def.type().data())
+    return registry_.Create(
+        OpKeyBuilder(operator_def.type().data())
+            .Device(type)
             .TypeConstraint("T", static_cast<DataType>(dtype))
             .Build(),
-                            operator_def,
-                            ws);
+        operator_def, ws);
   } else {
     return nullptr;
   }
 }
 
-OperatorBase::OperatorBase(const OperatorDef &operator_def, Workspace *ws)
-    : operator_ws_(ws),
-      operator_def_(std::make_shared<OperatorDef>(operator_def)) {}
+extern void Register_AddN(OperatorRegistry *op_registry);
+extern void Register_BatchNorm(OperatorRegistry *op_registry);
+extern void Register_BatchToSpaceND(OperatorRegistry *op_registry);
+extern void Register_BiasAdd(OperatorRegistry *op_registry);
+extern void Register_BufferToImage(OperatorRegistry *op_registry);
+extern void Register_ChannelShuffle(OperatorRegistry *op_registry);
+extern void Register_Concat(OperatorRegistry *op_registry);
+extern void Register_Conv2D(OperatorRegistry *op_registry);
+extern void Register_DepthwiseConv2d(OperatorRegistry *op_registry);
+extern void Register_FusedConv2D(OperatorRegistry *op_registry);
+extern void Register_GlobalAvgPooling(OperatorRegistry *op_registry);
+extern void Register_ImageToBuffer(OperatorRegistry *op_registry);
+extern void Register_Pooling(OperatorRegistry *op_registry);
+extern void Register_Relu(OperatorRegistry *op_registry);
+extern void Register_ResizeBilinear(OperatorRegistry *op_registry);
+extern void Register_SpaceToBatchND(OperatorRegistry *op_registry);
+
+OperatorRegistry::OperatorRegistry() {
+  Register_AddN(this);
+  Register_BatchNorm(this);
+  Register_BatchToSpaceND(this);
+  Register_BiasAdd(this);
+  Register_BufferToImage(this);
+  Register_ChannelShuffle(this);
+  Register_Concat(this);
+  Register_Conv2D(this);
+  Register_DepthwiseConv2d(this);
+  Register_FusedConv2D(this);
+  Register_GlobalAvgPooling(this);
+  Register_ImageToBuffer(this);
+  Register_Pooling(this);
+  Register_Relu(this);
+  Register_ResizeBilinear(this);
+  Register_SpaceToBatchND(this);
+}
 
 }  // namespace mace

mace/core/operator.h

@@ -5,13 +5,13 @@
 #ifndef MACE_CORE_OPERATOR_H
 #define MACE_CORE_OPERATOR_H
 
-#include "mace/core/common.h"
 #include "mace/core/arg_helper.h"
+#include "mace/core/common.h"
 #include "mace/core/future.h"
+#include "mace/core/public/mace.h"
 #include "mace/core/registry.h"
 #include "mace/core/tensor.h"
 #include "mace/core/workspace.h"
-#include "mace/core/public/mace.h"
 
 namespace mace {
 
@@ -122,29 +122,12 @@ class Operator : public OperatorBase {
 #define OP_OUTPUT_TAGS(first_input, ...) \
   enum _OutputTags { first_input = 0, __VA_ARGS__ }
 
-typedef Registry<std::string, OperatorBase, const OperatorDef &, Workspace *>
-    OperatorRegistry;
-typedef Registry<std::string, OperatorBase, const OperatorDef &, Workspace *> *(
-    *RegistryFunction)();
-std::map<int32_t, OperatorRegistry *> *gDeviceTypeRegistry();
-
-struct DeviceTypeRegisterer {
-  explicit DeviceTypeRegisterer(int32_t type, RegistryFunction func) {
-    if (gDeviceTypeRegistry()->count(type)) {
-      LOG(ERROR) << "Device type " << type
-                 << "registered twice. This should not happen. Did you have "
-                    "duplicated numbers assigned to different devices?";
-      std::exit(1);
-    }
-    // Calling the registry function to get the actual registry pointer.
-    gDeviceTypeRegistry()->emplace(type, func());
-  }
-};
-
 class OpKeyBuilder {
  public:
   explicit OpKeyBuilder(const char *op_name);
 
+  OpKeyBuilder &Device(DeviceType device);
+
   OpKeyBuilder &TypeConstraint(const char *attr_name, const DataType allowed);
 
   template <typename T>
@@ -154,6 +137,7 @@ class OpKeyBuilder {
 
  private:
   std::string op_name_;
+  DeviceType device_type_;
   std::map<std::string, DataType> type_constraint_;
 };
 
@@ -162,48 +146,30 @@ OpKeyBuilder &OpKeyBuilder::TypeConstraint(const char *attr_name) {
   return this->TypeConstraint(attr_name, DataTypeToEnum<T>::value);
 }
 
+class OperatorRegistry {
+ public:
+  typedef Registry<std::string, OperatorBase, const OperatorDef &, Workspace *>
+    RegistryType;
+  OperatorRegistry();
+  ~OperatorRegistry() = default;
+  RegistryType *registry() { return &registry_; };
+  std::unique_ptr<OperatorBase> CreateOperator(const OperatorDef &operator_def,
+                                               Workspace *ws,
+                                               DeviceType type,
+                                               const NetMode mode) const;
 
+ private:
+  RegistryType registry_;
+  DISABLE_COPY_AND_ASSIGN(OperatorRegistry);
+};
 
-#define MACE_REGISTER_DEVICE_TYPE(type, registry_function)         \
-  namespace {                                                      \
-  static DeviceTypeRegisterer MACE_ANONYMOUS_VARIABLE(DeviceType)( \
-      type, &registry_function);                                   \
-  }
-
-MACE_DECLARE_REGISTRY(CPUOperatorRegistry,
-                      OperatorBase,
-                      const OperatorDef &,
-                      Workspace *);
-
-#define REGISTER_CPU_OPERATOR_CREATOR(key, ...) \
-  MACE_REGISTER_CREATOR(CPUOperatorRegistry, key, __VA_ARGS__)
-#define REGISTER_CPU_OPERATOR(name, ...) \
-  MACE_REGISTER_CLASS(CPUOperatorRegistry, name, __VA_ARGS__)
-
-MACE_DECLARE_REGISTRY(NEONOperatorRegistry,
-                      OperatorBase,
-                      const OperatorDef &,
-                      Workspace *);
-
-#define REGISTER_NEON_OPERATOR_CREATOR(key, ...) \
-  MACE_REGISTER_CREATOR(NEONOperatorRegistry, key, __VA_ARGS__)
-#define REGISTER_NEON_OPERATOR(name, ...) \
-  MACE_REGISTER_CLASS(NEONOperatorRegistry, name, __VA_ARGS__)
-
-MACE_DECLARE_REGISTRY(OPENCLOperatorRegistry,
+MACE_DECLARE_REGISTRY(OpRegistry,
                       OperatorBase,
                       const OperatorDef &,
                       Workspace *);
 
-#define REGISTER_OPENCL_OPERATOR_CREATOR(key, ...) \
-  MACE_REGISTER_CREATOR(OPENCLOperatorRegistry, key, __VA_ARGS__)
-#define REGISTER_OPENCL_OPERATOR(name, ...) \
-  MACE_REGISTER_CLASS(OPENCLOperatorRegistry, name, __VA_ARGS__)
-
-unique_ptr<OperatorBase> CreateOperator(const OperatorDef &operator_def,
-                                        Workspace *ws,
-                                        DeviceType type,
-                                        const NetMode mode);
+#define REGISTER_OPERATOR(op_registry, name, ...) \
+  MACE_REGISTER_CLASS(OpRegistry, op_registry->registry(), name, __VA_ARGS__)
 
 }  //  namespace mace
 

mace/core/public/mace.h

@@ -302,10 +302,12 @@ class NetDef {
 
 class Workspace;
 class NetBase;
+class OperatorRegistry;
 
 class MaceEngine {
  public:
-  explicit MaceEngine(const NetDef *net_def, DeviceType device_type);
+  explicit MaceEngine(const NetDef *net_def,
+                      DeviceType device_type);
   ~MaceEngine();
   bool Run(const float *input,
            const std::vector<int64_t> &input_shape,
@@ -314,6 +316,7 @@ class MaceEngine {
   MaceEngine &operator=(const MaceEngine&) = delete;
 
  private:
+  std::shared_ptr<OperatorRegistry> op_registry_;
   DeviceType device_type_;
   std::unique_ptr<Workspace> ws_;
   std::unique_ptr<NetBase> net_;

mace/core/registry.h

@@ -17,24 +17,27 @@ class Registry {
   Registry() : registry_() {}
 
   void Register(const SrcType &key, Creator creator) {
+    VLOG(2) << "Registering: " << key;
     std::lock_guard<std::mutex> lock(register_mutex_);
     MACE_CHECK(registry_.count(key) == 0, "Key already registered.");
     registry_[key] = creator;
   }
 
-  inline bool Has(const SrcType &key) { return registry_.count(key) != 0; }
+  inline bool Has(const SrcType &key) const {
+    return registry_.count(key) != 0;
+  }
 
-  unique_ptr<ObjectType> Create(const SrcType &key, Args... args) {
+  unique_ptr<ObjectType> Create(const SrcType &key, Args... args) const {
     if (registry_.count(key) == 0) {
       LOG(FATAL) << "Key not registered: " << key;
     }
-    return registry_[key](args...);
+    return registry_.at(key)(args...);
   }
 
   /**
    * Returns the keys currently registered as a vector.
    */
-  vector<SrcType> Keys() {
+  vector<SrcType> Keys() const {
     vector<SrcType> keys;
     for (const auto &it : registry_) {
       keys.push_back(it.first);
@@ -77,39 +80,31 @@ class Registerer {
   typedef Registerer<SrcType, ObjectType, ##__VA_ARGS__>                    \
       Registerer##RegistryName;
 
+/*
 #define MACE_DEFINE_TYPED_REGISTRY(RegistryName, SrcType, ObjectType, ...) \
   Registry<SrcType, ObjectType, ##__VA_ARGS__> *RegistryName() {           \
     static Registry<SrcType, ObjectType, ##__VA_ARGS__> *registry =        \
         new Registry<SrcType, ObjectType, ##__VA_ARGS__>();                \
     return registry;                                                       \
   }
+*/
 
 #define MACE_DECLARE_REGISTRY(RegistryName, ObjectType, ...)         \
   MACE_DECLARE_TYPED_REGISTRY(RegistryName, std::string, ObjectType, \
                               ##__VA_ARGS__)
 
+/*
 #define MACE_DEFINE_REGISTRY(RegistryName, ObjectType, ...)         \
   MACE_DEFINE_TYPED_REGISTRY(RegistryName, std::string, ObjectType, \
                              ##__VA_ARGS__)
+*/
 
-#define MACE_REGISTER_TYPED_CREATOR(RegistryName, key, ...)                  \
-  namespace {                                                                \
-  static Registerer##RegistryName MACE_ANONYMOUS_VARIABLE(g_##RegistryName)( \
-      key, RegistryName(), __VA_ARGS__);
-
-#define MACE_REGISTER_TYPED_CLASS(RegistryName, key, ...)                    \
-  namespace {                                                                \
-  static Registerer##RegistryName MACE_ANONYMOUS_VARIABLE(g_##RegistryName)( \
-      key,                                                                   \
-      RegistryName(),                                                        \
-      Registerer##RegistryName::DefaultCreator<__VA_ARGS__>);                \
-  }
-
-#define MACE_REGISTER_CREATOR(RegistryName, key, ...) \
-  MACE_REGISTER_TYPED_CREATOR(RegistryName, key, __VA_ARGS__)
+#define MACE_REGISTER_TYPED_CLASS(RegistryName, registry, key, ...)   \
+  Registerer##RegistryName MACE_ANONYMOUS_VARIABLE(l_##RegistryName)( \
+      key, registry, Registerer##RegistryName::DefaultCreator<__VA_ARGS__>);
 
-#define MACE_REGISTER_CLASS(RegistryName, key, ...) \
-  MACE_REGISTER_TYPED_CLASS(RegistryName, key, __VA_ARGS__)
+#define MACE_REGISTER_CLASS(RegistryName, registry, key, ...) \
+  MACE_REGISTER_TYPED_CLASS(RegistryName, registry, key, __VA_ARGS__)
 
 }  // namespace mace
 

mace/core/runtime/opencl/opencl_allocator.cc

@@ -127,6 +127,4 @@ void OpenCLAllocator::Unmap(void *buffer, void *mapped_ptr) {
 
 bool OpenCLAllocator::OnHost() { return false; }
 
-MACE_REGISTER_ALLOCATOR(DeviceType::OPENCL, new OpenCLAllocator());
-
 }  // namespace mace

mace/core/tensor.h

@@ -105,7 +105,8 @@ class Tensor {
   inline index_t dim_size() const { return shape_.size(); }
 
   inline index_t dim(unsigned int index) const {
-    MACE_CHECK(index < shape_.size(), "Exceeding ndim limit");
+    MACE_CHECK(index < shape_.size(), "Dim out of range: ",
+                                      index, " >= ", shape_.size());
     return shape_[index];
   }
 

mace/examples/BUILD

@@ -11,7 +11,6 @@ cc_binary(
     deps = [
         "//mace/core",
         "//mace/ops",
-        "//mace/core:opencl_runtime",
     ],
 )
 

mace/kernels/BUILD

@@ -26,7 +26,6 @@ cc_library(
     linkopts = if_android(["-lm"]),
     deps = [
         "//mace/core",
-        "//mace/core:opencl_runtime",
         "//mace/utils:utils_hdrs",
     ],
 )

mace/ops/addn.cc

@@ -6,26 +6,32 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("AddN")
+void Register_AddN(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("AddN")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     AddNOp<DeviceType::CPU, float>);
 
 #if MACE_ENABLE_NEON
-REGISTER_NEON_OPERATOR(OpKeyBuilder("AddN")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("AddN")
+                                     .Device(DeviceType::NEON)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     AddNOp<DeviceType::NEON, float>);
 #endif  // MACE_ENABLE_NEON
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("AddN")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("AddN")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     AddNOp<DeviceType::OPENCL, float>);
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("AddN")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("AddN")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     AddNOp<DeviceType::OPENCL, half>);
+}
 
 }  //  namespace mace

mace/ops/addn_benchmark.cc

@@ -15,8 +15,8 @@ static void AddNBenchmark(int iters, int inputs, int n, int h, int w, int c) {
   OpsTestNet net;
   // Add input data
   for (int i = 0; i < inputs; ++i) {
-    net.AddRandomInput<D, float>(
-        internal::MakeString("Input", i).c_str(), {n, h, w, c});
+    net.AddRandomInput<D, float>(internal::MakeString("Input", i).c_str(),
+                                 {n, h, w, c});
   }
 
   if (D == DeviceType::OPENCL) {

mace/ops/batch_norm.cc

@@ -6,26 +6,32 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("BatchNorm")
+void Register_BatchNorm(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("BatchNorm")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     BatchNormOp<DeviceType::CPU, float>);
 
 #if MACE_ENABLE_NEON
-REGISTER_NEON_OPERATOR(OpKeyBuilder("BatchNorm")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("BatchNorm")
+                                     .Device(DeviceType::NEON)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     BatchNormOp<DeviceType::NEON, float>);
 #endif  // MACE_ENABLE_NEON
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("BatchNorm")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("BatchNorm")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     BatchNormOp<DeviceType::OPENCL, float>);
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("BatchNorm")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("BatchNorm")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     BatchNormOp<DeviceType::OPENCL, half>);
+}
 
 }  //  namespace mace

mace/ops/batch_norm_benchmark.cc

@@ -23,11 +23,16 @@ static void BatchNorm(
   net.AddRandomInput<D, T>("Var", {channels}, true);
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, float>(net, "Scale", "ScaleImage", kernels::BufferType::ARGUMENT);
-    BufferToImage<D, float>(net, "Offset", "OffsetImage", kernels::BufferType::ARGUMENT);
-    BufferToImage<D, float>(net, "Mean", "MeanImage", kernels::BufferType::ARGUMENT);
-    BufferToImage<D, float>(net, "Var", "VarImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, float>(net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT);
+    BufferToImage<D, float>(net, "Scale", "ScaleImage",
+                            kernels::BufferType::ARGUMENT);
+    BufferToImage<D, float>(net, "Offset", "OffsetImage",
+                            kernels::BufferType::ARGUMENT);
+    BufferToImage<D, float>(net, "Mean", "MeanImage",
+                            kernels::BufferType::ARGUMENT);
+    BufferToImage<D, float>(net, "Var", "VarImage",
+                            kernels::BufferType::ARGUMENT);
     OpDefBuilder("BatchNorm", "BatchNormBM")
         .Input("InputImage")
         .Input("ScaleImage")
@@ -37,8 +42,7 @@ static void BatchNorm(
         .AddFloatArg("epsilon", 1e-3)
         .Output("Output")
         .Finalize(net.NewOperatorDef());
-  }
-  else {
+  } else {
     OpDefBuilder("BatchNorm", "BatchNormBM")
         .Input("Input")
         .Input("Scale")
@@ -50,7 +54,6 @@ static void BatchNorm(
         .Finalize(net.NewOperatorDef());
   }
 
-
   // tuning
   setenv("MACE_TUNING", "1", 1);
   net.RunOp(D);
@@ -81,7 +84,6 @@ static void BatchNorm(
 
 #define BM_BATCH_NORM(N, C, H, W, TYPE)        \
   BM_BATCH_NORM_MACRO(N, C, H, W, TYPE, CPU);  \
-  BM_BATCH_NORM_MACRO(N, C, H, W, TYPE, NEON);\
   BM_BATCH_NORM_MACRO(N, C, H, W, TYPE, OPENCL);
 
 BM_BATCH_NORM(1, 1, 512, 512, float);

mace/ops/batch_norm_test.cc

@@ -22,11 +22,16 @@ void Simple() {
   net.AddInputFromArray<D, float>("Var", {1}, {11.67f});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, float>(net, "Scale", "ScaleImage", kernels::BufferType::ARGUMENT);
-    BufferToImage<D, float>(net, "Offset", "OffsetImage", kernels::BufferType::ARGUMENT);
-    BufferToImage<D, float>(net, "Mean", "MeanImage", kernels::BufferType::ARGUMENT);
-    BufferToImage<D, float>(net, "Var", "VarImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, float>(net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT);
+    BufferToImage<D, float>(net, "Scale", "ScaleImage",
+                            kernels::BufferType::ARGUMENT);
+    BufferToImage<D, float>(net, "Offset", "OffsetImage",
+                            kernels::BufferType::ARGUMENT);
+    BufferToImage<D, float>(net, "Mean", "MeanImage",
+                            kernels::BufferType::ARGUMENT);
+    BufferToImage<D, float>(net, "Var", "VarImage",
+                            kernels::BufferType::ARGUMENT);
 
     OpDefBuilder("BatchNorm", "BatchNormTest")
         .Input("InputImage")
@@ -41,7 +46,8 @@ void Simple() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT);
   } else {
     OpDefBuilder("BatchNorm", "BatchNormTest")
         .Input("Input")
@@ -64,9 +70,7 @@ void Simple() {
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-2);
 }
 
-TEST_F(BatchNormOpTest, SimpleCPU) {
-  Simple<DeviceType::CPU>();
-}
+TEST_F(BatchNormOpTest, SimpleCPU) { Simple<DeviceType::CPU>(); }
 
 /*
 TEST_F(BatchNormOpTest, SimpleNEON) {
@@ -74,9 +78,7 @@ TEST_F(BatchNormOpTest, SimpleNEON) {
 }
 */
 
-TEST_F(BatchNormOpTest, SimpleOPENCL) {
-  Simple<DeviceType::OPENCL>();
-}
+TEST_F(BatchNormOpTest, SimpleOPENCL) { Simple<DeviceType::OPENCL>(); }
 
 /*
 TEST_F(BatchNormOpTest, SimpleRandomNeon) {
@@ -100,7 +102,8 @@ TEST_F(BatchNormOpTest, SimpleRandomNeon) {
       .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<DeviceType::CPU, float>("Input", {batch, channels, height, width});
+  net.AddRandomInput<DeviceType::CPU, float>("Input", {batch, channels, height,
+width});
   net.AddRandomInput<DeviceType::CPU, float>("Scale", {channels});
   net.AddRandomInput<DeviceType::CPU, float>("Offset", {channels});
   net.AddRandomInput<DeviceType::CPU, float>("Mean", {channels});
@@ -141,7 +144,8 @@ TEST_F(BatchNormOpTest, ComplexRandomNeon) {
       .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<DeviceType::CPU, float>("Input", {batch, channels, height, width});
+  net.AddRandomInput<DeviceType::CPU, float>("Input", {batch, channels, height,
+width});
   net.AddRandomInput<DeviceType::CPU, float>("Scale", {channels});
   net.AddRandomInput<DeviceType::CPU, float>("Offset", {channels});
   net.AddRandomInput<DeviceType::CPU, float>("Mean", {channels});
@@ -184,7 +188,8 @@ TEST_F(BatchNormOpTest, SimpleRandomOPENCL) {
       .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<DeviceType::OPENCL, float>("Input", {batch, height, width, channels});
+  net.AddRandomInput<DeviceType::OPENCL, float>(
+      "Input", {batch, height, width, channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Scale", {channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Offset", {channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Mean", {channels});
@@ -198,11 +203,16 @@ TEST_F(BatchNormOpTest, SimpleRandomOPENCL) {
   expected.Copy(*net.GetOutput("Output"));
 
   // Run on opencl
-  BufferToImage<DeviceType::OPENCL, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-  BufferToImage<DeviceType::OPENCL, float>(net, "Scale", "ScaleImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, float>(net, "Offset", "OffsetImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, float>(net, "Mean", "MeanImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, float>(net, "Var", "VarImage", kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Input", "InputImage",
+                                           kernels::BufferType::IN_OUT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Scale", "ScaleImage",
+                                           kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Offset", "OffsetImage",
+                                           kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Mean", "MeanImage",
+                                           kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Var", "VarImage",
+                                           kernels::BufferType::ARGUMENT);
 
   OpDefBuilder("BatchNorm", "BatchNormTest")
       .Input("InputImage")
@@ -223,7 +233,8 @@ TEST_F(BatchNormOpTest, SimpleRandomOPENCL) {
   net.RunOp(DeviceType::OPENCL);
   net.Sync();
 
-  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput",
+                                           kernels::BufferType::IN_OUT);
   ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2);
 }
 
@@ -249,7 +260,8 @@ TEST_F(BatchNormOpTest, SimpleRandomHalfOPENCL) {
       .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<DeviceType::OPENCL, float>("Input", {batch, height, width, channels});
+  net.AddRandomInput<DeviceType::OPENCL, float>(
+      "Input", {batch, height, width, channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Scale", {channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Offset", {channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Mean", {channels});
@@ -263,11 +275,16 @@ TEST_F(BatchNormOpTest, SimpleRandomHalfOPENCL) {
   expected.Copy(*net.GetOutput("Output"));
 
   // Run on opencl
-  BufferToImage<DeviceType::OPENCL, half>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-  BufferToImage<DeviceType::OPENCL, half>(net, "Scale", "ScaleImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, half>(net, "Offset", "OffsetImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, half>(net, "Mean", "MeanImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, half>(net, "Var", "VarImage", kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, half>(net, "Input", "InputImage",
+                                          kernels::BufferType::IN_OUT);
+  BufferToImage<DeviceType::OPENCL, half>(net, "Scale", "ScaleImage",
+                                          kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, half>(net, "Offset", "OffsetImage",
+                                          kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, half>(net, "Mean", "MeanImage",
+                                          kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, half>(net, "Var", "VarImage",
+                                          kernels::BufferType::ARGUMENT);
 
   OpDefBuilder("BatchNorm", "BatchNormTest")
       .Input("InputImage")
@@ -289,7 +306,8 @@ TEST_F(BatchNormOpTest, SimpleRandomHalfOPENCL) {
   net.RunOp(DeviceType::OPENCL);
   net.Sync();
 
-  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput",
+                                           kernels::BufferType::IN_OUT);
   ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.5);
 }
 
@@ -315,7 +333,8 @@ TEST_F(BatchNormOpTest, ComplexRandomOPENCL) {
       .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<DeviceType::OPENCL, float>("Input", {batch, height, width, channels});
+  net.AddRandomInput<DeviceType::OPENCL, float>(
+      "Input", {batch, height, width, channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Scale", {channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Offset", {channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Mean", {channels});
@@ -328,13 +347,17 @@ TEST_F(BatchNormOpTest, ComplexRandomOPENCL) {
   Tensor expected;
   expected.Copy(*net.GetOutput("Output"));
 
-
   // Run on opencl
-  BufferToImage<DeviceType::OPENCL, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-  BufferToImage<DeviceType::OPENCL, float>(net, "Scale", "ScaleImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, float>(net, "Offset", "OffsetImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, float>(net, "Mean", "MeanImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, float>(net, "Var", "VarImage", kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Input", "InputImage",
+                                           kernels::BufferType::IN_OUT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Scale", "ScaleImage",
+                                           kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Offset", "OffsetImage",
+                                           kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Mean", "MeanImage",
+                                           kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Var", "VarImage",
+                                           kernels::BufferType::ARGUMENT);
 
   OpDefBuilder("BatchNorm", "BatchNormTest")
       .Input("InputImage")
@@ -355,7 +378,8 @@ TEST_F(BatchNormOpTest, ComplexRandomOPENCL) {
   net.RunOp(DeviceType::OPENCL);
   net.Sync();
 
-  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput",
+                                           kernels::BufferType::IN_OUT);
   ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2);
 }
 
@@ -381,7 +405,8 @@ TEST_F(BatchNormOpTest, ComplexRandomHalfOPENCL) {
       .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<DeviceType::OPENCL, float>("Input", {batch, height, width, channels});
+  net.AddRandomInput<DeviceType::OPENCL, float>(
+      "Input", {batch, height, width, channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Scale", {channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Offset", {channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Mean", {channels});
@@ -394,13 +419,17 @@ TEST_F(BatchNormOpTest, ComplexRandomHalfOPENCL) {
   Tensor expected;
   expected.Copy(*net.GetOutput("Output"));
 
-
   // Run on opencl
-  BufferToImage<DeviceType::OPENCL, half>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-  BufferToImage<DeviceType::OPENCL, half>(net, "Scale", "ScaleImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, half>(net, "Offset", "OffsetImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, half>(net, "Mean", "MeanImage", kernels::BufferType::ARGUMENT);
-  BufferToImage<DeviceType::OPENCL, half>(net, "Var", "VarImage", kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, half>(net, "Input", "InputImage",
+                                          kernels::BufferType::IN_OUT);
+  BufferToImage<DeviceType::OPENCL, half>(net, "Scale", "ScaleImage",
+                                          kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, half>(net, "Offset", "OffsetImage",
+                                          kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, half>(net, "Mean", "MeanImage",
+                                          kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, half>(net, "Var", "VarImage",
+                                          kernels::BufferType::ARGUMENT);
 
   OpDefBuilder("BatchNorm", "BatchNormTest")
       .Input("InputImage")
@@ -422,7 +451,8 @@ TEST_F(BatchNormOpTest, ComplexRandomHalfOPENCL) {
   net.RunOp(DeviceType::OPENCL);
   net.Sync();
 
-  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput",
+                                           kernels::BufferType::IN_OUT);
   ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.5);
 }
 }

mace/ops/batch_to_space.cc

@@ -6,13 +6,17 @@
 
 namespace mace {
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("BatchToSpaceND")
+void Register_BatchToSpaceND(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("BatchToSpaceND")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     BatchToSpaceNDOp<DeviceType::OPENCL, float>);
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("BatchToSpaceND")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("BatchToSpaceND")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     BatchToSpaceNDOp<DeviceType::OPENCL, half>);
+}
 
 }  //  namespace mace

mace/ops/batch_to_space_benchmark.cc

@@ -14,7 +14,8 @@ static void BMBatchToSpace(
 
   OpsTestNet net;
   net.AddRandomInput<D, float>("Input", {batch, height, width, channels});
-  BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+  BufferToImage<D, float>(net, "Input", "InputImage",
+                          kernels::BufferType::IN_OUT);
   OpDefBuilder("BatchToSpaceND", "BatchToSpaceNDTest")
       .Input("InputImage")
       .Output("OutputImage")
@@ -36,7 +37,8 @@ static void BMBatchToSpace(
 }
 
 #define BM_BATCH_TO_SPACE_MACRO(N, H, W, C, ARG, TYPE, DEVICE)             \
-  static void BM_BATCH_TO_SPACE_##N##_##H##_##W##_##C##_##ARG##_##TYPE##_##DEVICE( \
+  static void                                                              \
+      BM_BATCH_TO_SPACE_##N##_##H##_##W##_##C##_##ARG##_##TYPE##_##DEVICE( \
           int iters) {                                                     \
     const int64_t tot = static_cast<int64_t>(iters) * N * C * H * W;       \
     mace::testing::ItemsProcessed(tot);                                    \

mace/ops/bias_add.cc

@@ -6,28 +6,34 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("BiasAdd")
+void Register_BiasAdd(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("BiasAdd")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     BiasAddOp<DeviceType::CPU, float>);
 
-/*
-#if __ARM_NEON
-REGISTER_NEON_OPERATOR(OpKeyBuilder("BiasAdd")
+  /*
+  #if __ARM_NEON
+  REGISTER_OPERATOR(op_registry,OpKeyBuilder("BiasAdd")
+                               .Device(DeviceType::NEON)
                                .TypeConstraint<float>("T")
                                .Build(),
                          BiasAddOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
-*/
+  #endif  // __ARM_NEON
+  */
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("BiasAdd")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("BiasAdd")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     BiasAddOp<DeviceType::OPENCL, float>);
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("BiasAdd")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("BiasAdd")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     BiasAddOp<DeviceType::OPENCL, half>);
+}
 
 }  //  namespace mace

mace/ops/bias_add_benchmark.cc

@@ -9,8 +9,7 @@
 
 namespace mace {
 template <DeviceType D, typename T>
-static void BiasAdd(
-    int iters, int batch, int channels, int height, int width) {
+static void BiasAdd(int iters, int batch, int channels, int height, int width) {
   mace::testing::StopTiming();
 
   OpsTestNet net;
@@ -20,15 +19,16 @@ static void BiasAdd(
   net.AddRandomInput<D, T>("Bias", {channels}, true);
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Bias", "BiasImage",
+                        kernels::BufferType::ARGUMENT);
     OpDefBuilder("BiasAdd", "BiasAddBM")
         .Input("InputImage")
         .Input("BiasImage")
         .Output("Output")
         .Finalize(net.NewOperatorDef());
-  }
-  else {
+  } else {
     OpDefBuilder("BiasAdd", "BiasAddBM")
         .Input("Input")
         .Input("Bias")

mace/ops/bias_add_test.cc

@@ -19,8 +19,10 @@ void BiasAddSimple() {
   net.AddInputFromArray<D, float>("Bias", {1}, {0.5f});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, float>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, float>(net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT);
+    BufferToImage<D, float>(net, "Bias", "BiasImage",
+                            kernels::BufferType::ARGUMENT);
 
     OpDefBuilder("BiasAdd", "BiasAddTest")
         .Input("InputImage")
@@ -31,7 +33,8 @@ void BiasAddSimple() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT);
   } else {
     OpDefBuilder("BiasAdd", "BiasAddTest")
         .Input("Input")
@@ -43,16 +46,14 @@ void BiasAddSimple() {
   }
 
   // Check
-  auto expected =
-      CreateTensor<float>({1, 6, 2, 1}, {5.5, 5.5, 7.5, 7.5, 9.5, 9.5, 11.5,
-                                         11.5, 13.5, 13.5, 15.5, 15.5});
+  auto expected = CreateTensor<float>(
+      {1, 6, 2, 1},
+      {5.5, 5.5, 7.5, 7.5, 9.5, 9.5, 11.5, 11.5, 13.5, 13.5, 15.5, 15.5});
 
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-2);
 }
 
-TEST_F(BiasAddOpTest, BiasAddSimpleCPU) {
-  BiasAddSimple<DeviceType::CPU>();
-}
+TEST_F(BiasAddOpTest, BiasAddSimpleCPU) { BiasAddSimple<DeviceType::CPU>(); }
 
 TEST_F(BiasAddOpTest, BiasAddSimpleOPENCL) {
   BiasAddSimple<DeviceType::OPENCL>();
@@ -76,7 +77,8 @@ TEST_F(BiasAddOpTest, SimpleRandomOPENCL) {
       .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<DeviceType::OPENCL, float>("Input", {batch, height, width, channels});
+  net.AddRandomInput<DeviceType::OPENCL, float>(
+      "Input", {batch, height, width, channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Bias", {channels}, true);
 
   // run cpu
@@ -87,8 +89,10 @@ TEST_F(BiasAddOpTest, SimpleRandomOPENCL) {
   expected.Copy(*net.GetOutput("Output"));
 
   // Run on opencl
-  BufferToImage<DeviceType::OPENCL, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-  BufferToImage<DeviceType::OPENCL, float>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Input", "InputImage",
+                                           kernels::BufferType::IN_OUT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Bias", "BiasImage",
+                                           kernels::BufferType::ARGUMENT);
 
   OpDefBuilder("BiasAdd", "BiasAddTest")
       .Input("InputImage")
@@ -100,7 +104,8 @@ TEST_F(BiasAddOpTest, SimpleRandomOPENCL) {
   net.RunOp(DeviceType::OPENCL);
   net.Sync();
 
-  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput",
+                                           kernels::BufferType::IN_OUT);
   ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2);
 }
 
@@ -122,7 +127,8 @@ TEST_F(BiasAddOpTest, ComplexRandomOPENCL) {
       .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<DeviceType::OPENCL, float>("Input", {batch, height, width, channels});
+  net.AddRandomInput<DeviceType::OPENCL, float>(
+      "Input", {batch, height, width, channels});
   net.AddRandomInput<DeviceType::OPENCL, float>("Bias", {channels}, true);
 
   // run cpu
@@ -132,10 +138,11 @@ TEST_F(BiasAddOpTest, ComplexRandomOPENCL) {
   Tensor expected;
   expected.Copy(*net.GetOutput("Output"));
 
-
   // Run on opencl
-  BufferToImage<DeviceType::OPENCL, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-  BufferToImage<DeviceType::OPENCL, float>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Input", "InputImage",
+                                           kernels::BufferType::IN_OUT);
+  BufferToImage<DeviceType::OPENCL, float>(net, "Bias", "BiasImage",
+                                           kernels::BufferType::ARGUMENT);
 
   OpDefBuilder("BiasAdd", "BiasAddTest")
       .Input("InputImage")
@@ -147,8 +154,8 @@ TEST_F(BiasAddOpTest, ComplexRandomOPENCL) {
   net.RunOp(DeviceType::OPENCL);
   net.Sync();
 
-  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "OPENCLOutput",
+                                           kernels::BufferType::IN_OUT);
   ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 1e-2);
 }
-
 }

mace/ops/buffer_to_image.cc

@@ -6,14 +6,18 @@
 
 namespace mace {
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("BufferToImage")
+void Register_BufferToImage(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("BufferToImage")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     BufferToImageOp<DeviceType::OPENCL, float>);
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("BufferToImage")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("BufferToImage")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     BufferToImageOp<DeviceType::OPENCL, half>);
+}
 
 }  //  namespace mace

mace/ops/channel_shuffle.cc

@@ -6,9 +6,12 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("ChannelShuffle")
+void Register_ChannelShuffle(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("ChannelShuffle")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     ChannelShuffleOp<DeviceType::CPU, float>);
+}
 
 }  // namespace mace

mace/ops/channel_shuffle_benchmark.cc

@@ -23,7 +23,8 @@ static void ChannelShuffle(
       .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<DeviceType::CPU, float>("Input", {batch, channels, height, width});
+  net.AddRandomInput<DeviceType::CPU, float>("Input",
+                                             {batch, channels, height, width});
 
   // Warm-up
   for (int i = 0; i < 5; ++i) {

mace/ops/channel_shuffle_test.cc

@@ -17,7 +17,6 @@ TEST_F(ChannelShuffleOpTest, C8G4) {
       .AddIntArg("group", 4)
       .Finalize(net.NewOperatorDef());
 
-
   // Add input data
   net.AddInputFromArray<DeviceType::CPU, float>(
       "Input", {1, 8, 1, 2},

mace/ops/concat.cc

@@ -6,21 +6,28 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("Concat")
+void Register_Concat(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Concat")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     ConcatOp<DeviceType::CPU, float>);
-REGISTER_CPU_OPERATOR(OpKeyBuilder("Concat")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Concat")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     ConcatOp<DeviceType::CPU, half>);
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Concat")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Concat")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     ConcatOp<DeviceType::OPENCL, float>);
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Concat")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Concat")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     ConcatOp<DeviceType::OPENCL, half>);
+}
+
 }  // namespace mace

mace/ops/concat_benchmark.cc

@@ -60,8 +60,10 @@ static void OpenclConcatHelper(int iters,
   net.AddRandomInput<DeviceType::OPENCL, float>("Input0", shape0);
   net.AddRandomInput<DeviceType::OPENCL, float>("Input1", shape1);
 
-  BufferToImage<DeviceType::OPENCL, T>(net, "Input0", "InputImage0", kernels::BufferType::IN_OUT);
-  BufferToImage<DeviceType::OPENCL, T>(net, "Input1", "InputImage1", kernels::BufferType::IN_OUT);
+  BufferToImage<DeviceType::OPENCL, T>(net, "Input0", "InputImage0",
+                                       kernels::BufferType::IN_OUT);
+  BufferToImage<DeviceType::OPENCL, T>(net, "Input1", "InputImage1",
+                                       kernels::BufferType::IN_OUT);
   OpDefBuilder("Concat", "ConcatBM")
       .Input("InputImage0")
       .Input("InputImage1")
@@ -75,7 +77,8 @@ static void OpenclConcatHelper(int iters,
     net.RunOp(DeviceType::OPENCL);
   }
 
-  const int64_t tot = static_cast<int64_t>(iters) *
+  const int64_t tot =
+      static_cast<int64_t>(iters) *
       (net.GetTensor("Input0")->size() + net.GetTensor("Input1")->size());
   mace::testing::ItemsProcessed(tot);
   testing::BytesProcessed(tot * sizeof(T));

mace/ops/concat_test.cc

@@ -97,7 +97,9 @@ TEST_F(ConcatOpTest, CPURandom) {
   for (int i = 0; i < num_inputs; ++i) {
     builder = builder.Input(("Input" + ToString(i)).c_str());
   }
-  builder.AddIntArg("axis", axis).Output("Output").Finalize(net.NewOperatorDef());
+  builder.AddIntArg("axis", axis)
+      .Output("Output")
+      .Finalize(net.NewOperatorDef());
 
   std::vector<index_t> shape_data;
   GenerateRandomIntTypeData<index_t>({dim}, shape_data, 1, dim);
@@ -110,8 +112,8 @@ TEST_F(ConcatOpTest, CPURandom) {
     concat_axis_size += input_shapes[i][axis];
     GenerateRandomRealTypeData(input_shapes[i], inputs[i]);
     input_ptrs[i] = inputs[i].data();
-    net.AddInputFromArray<DeviceType::CPU, float>(("Input" + ToString(i)).c_str(),
-                                 input_shapes[i], inputs[i]);
+    net.AddInputFromArray<DeviceType::CPU, float>(
+        ("Input" + ToString(i)).c_str(), input_shapes[i], inputs[i]);
   }
 
   // Run
@@ -137,7 +139,7 @@ TEST_F(ConcatOpTest, CPURandom) {
   }
 }
 
-template<typename T>
+template <typename T>
 void OpenclRandomTest(const std::vector<std::vector<index_t>> &shapes,
                       const int axis) {
   srand(time(nullptr));
@@ -149,9 +151,9 @@ void OpenclRandomTest(const std::vector<std::vector<index_t>> &shapes,
     const std::string input_name = ("Input" + ToString(i)).c_str();
     const std::string image_name = ("InputImage" + ToString(i)).c_str();
     concat_axis_size += shapes[i][axis];
-    net.AddRandomInput<DeviceType::OPENCL, float>(input_name,
-                                                  shapes[i]);
-    BufferToImage<DeviceType::OPENCL, T>(net, input_name, image_name, kernels::BufferType::IN_OUT);
+    net.AddRandomInput<DeviceType::OPENCL, float>(input_name, shapes[i]);
+    BufferToImage<DeviceType::OPENCL, T>(net, input_name, image_name,
+                                         kernels::BufferType::IN_OUT);
   }
 
   auto builder = OpDefBuilder("Concat", "ConcatTest");
@@ -167,7 +169,8 @@ void OpenclRandomTest(const std::vector<std::vector<index_t>> &shapes,
   // Run
   net.RunOp(DeviceType::OPENCL);
 
-  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+  ImageToBuffer<DeviceType::OPENCL, float>(net, "OutputImage", "Output",
+                                           kernels::BufferType::IN_OUT);
 
   // Check
   auto output = net.GetOutput("Output");
@@ -182,15 +185,16 @@ void OpenclRandomTest(const std::vector<std::vector<index_t>> &shapes,
   while (output_ptr != (output->data<float>() + output->size())) {
     for (int i = 0; i < num_inputs; ++i) {
       index_t num_elements =
-          std::accumulate(shapes[i].begin() + axis, shapes[i].end(),
-                          1, std::multiplies<index_t>());
+          std::accumulate(shapes[i].begin() + axis, shapes[i].end(), 1,
+                          std::multiplies<index_t>());
 
       const std::string input_name = ("Input" + ToString(i)).c_str();
       const Tensor *input_tensor = net.GetTensor(input_name.data());
       Tensor::MappingGuard input_guard(input_tensor);
       const float *input_ptr = input_tensor->data<float>() + k * num_elements;
       for (int j = 0; j < num_elements; ++j) {
-        EXPECT_NEAR(*(input_ptr + j), *output_ptr++, 1e-2) << "With index: " << i << ", " << j;
+        EXPECT_NEAR(*(input_ptr + j), *output_ptr++, 1e-2)
+            << "With index: " << i << ", " << j;
       }
     }
     k++;
@@ -198,25 +202,13 @@ void OpenclRandomTest(const std::vector<std::vector<index_t>> &shapes,
 }
 
 TEST_F(ConcatOpTest, OPENCLAligned) {
-  OpenclRandomTest<float>({
-                              {3, 32, 32, 32},
-                              {3, 32, 32, 64}
-                          },
-                          3);
+  OpenclRandomTest<float>({{3, 32, 32, 32}, {3, 32, 32, 64}}, 3);
 }
 
 TEST_F(ConcatOpTest, OPENCLHalfAligned) {
-  OpenclRandomTest<half>({
-                              {3, 32, 32, 32},
-                              {3, 32, 32, 64}
-                          },
-                          3);
+  OpenclRandomTest<half>({{3, 32, 32, 32}, {3, 32, 32, 64}}, 3);
 }
 
 TEST_F(ConcatOpTest, OPENCLUnAligned) {
-  OpenclRandomTest<float>({
-                              {3, 32, 32, 13},
-                              {3, 32, 32, 17}
-                          },
-                          3);
+  OpenclRandomTest<float>({{3, 32, 32, 13}, {3, 32, 32, 17}}, 3);
 }

mace/ops/conv_2d.cc

@@ -6,31 +6,38 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("Conv2D")
+void Register_Conv2D(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Conv2D")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     Conv2dOp<DeviceType::CPU, float>);
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("Conv2D")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Conv2D")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     Conv2dOp<DeviceType::CPU, half>);
 
 #if MACE_ENABLE_NEON
-REGISTER_NEON_OPERATOR(OpKeyBuilder("Conv2D")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Conv2D")
+                                     .Device(DeviceType::NEON)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     Conv2dOp<DeviceType::NEON, float>);
 #endif  // MACE_ENABLE_NEON
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Conv2D")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Conv2D")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     Conv2dOp<DeviceType::OPENCL, float>);
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Conv2D")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Conv2D")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     Conv2dOp<DeviceType::OPENCL, half>);
+}
 
 }  // namespace mace

mace/ops/conv_2d_benchmark.cc

@@ -33,9 +33,12 @@ static void Conv2d(int iters,
   net.AddRandomInput<D, float>("Bias", {output_channels});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, T>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Filter", "FilterImage",
+                        kernels::BufferType::FILTER);
+    BufferToImage<D, T>(net, "Bias", "BiasImage",
+                        kernels::BufferType::ARGUMENT);
     OpDefBuilder("Conv2D", "Conv2dTest")
         .Input("InputImage")
         .Input("FilterImage")
@@ -106,28 +109,29 @@ BM_CONV_2D(1, 3, 512, 512, 7, 7, 2, SAME, 64, half);
 BM_CONV_2D(1, 512, 64, 64, 1, 1, 1, SAME, 256, half);
 
 // Test RGB <-> YUV
-//BM_CONV_2D(1, 3, 2160, 1080, 1, 1, 1, VALID, 3, float);
-//BM_CONV_2D(1, 3, 480, 480, 1, 1, 1, VALID, 3, float);
+// BM_CONV_2D(1, 3, 2160, 1080, 1, 1, 1, VALID, 3, float);
+// BM_CONV_2D(1, 3, 480, 480, 1, 1, 1, VALID, 3, float);
 //
-//BM_CONV_2D(1, 64, 32, 32, 1, 1, 1, VALID, 128, float);
-//BM_CONV_2D(1, 64, 33, 31, 1, 1, 1, VALID, 128, float);  // Test bad alignments
-//BM_CONV_2D(1, 3, 512, 512, 1, 1, 1, VALID, 3, float);
-//BM_CONV_2D(1, 32, 112, 112, 1, 1, 1, VALID, 64, float);
-//BM_CONV_2D(1, 64, 56, 56, 1, 1, 1, VALID, 128, float);
-//BM_CONV_2D(1, 256, 28, 28, 1, 1, 1, VALID, 256, float);
-//BM_CONV_2D(1, 1024, 7, 7, 1, 1, 1, VALID, 1024, float);
-//BM_CONV_2D(1, 64, 32, 32, 3, 3, 1, VALID, 128, float);
-//BM_CONV_2D(1, 64, 33, 31, 3, 3, 1, VALID, 128, float);
-//BM_CONV_2D(1, 3, 512, 512, 3, 3, 1, VALID, 3, float);
-//BM_CONV_2D(1, 64, 32, 32, 3, 3, 1, SAME, 128, float);
-//BM_CONV_2D(1, 64, 33, 31, 3, 3, 1, SAME, 128, float);
-//BM_CONV_2D(1, 64, 32, 32, 3, 3, 2, VALID, 128, float);
-//BM_CONV_2D(1, 3, 512, 512, 3, 3, 2, VALID, 3, float);
-//BM_CONV_2D(1, 64, 33, 31, 3, 3, 2, VALID, 128, float);
-//BM_CONV_2D(1, 64, 32, 32, 3, 3, 2, SAME, 128, float);
-//BM_CONV_2D(1, 64, 33, 31, 3, 3, 2, SAME, 128, float);
-//BM_CONV_2D(1, 64, 32, 32, 5, 5, 1, VALID, 128, float);
-//BM_CONV_2D(1, 64, 32, 31, 5, 5, 1, VALID, 128, float);
-//BM_CONV_2D(1, 64, 32, 32, 5, 5, 1, SAME, 128, float);
-//BM_CONV_2D(1, 64, 32, 31, 5, 5, 1, SAME, 128, float);
+// BM_CONV_2D(1, 64, 32, 32, 1, 1, 1, VALID, 128, float);
+// BM_CONV_2D(1, 64, 33, 31, 1, 1, 1, VALID, 128, float);  // Test bad
+// alignments
+// BM_CONV_2D(1, 3, 512, 512, 1, 1, 1, VALID, 3, float);
+// BM_CONV_2D(1, 32, 112, 112, 1, 1, 1, VALID, 64, float);
+// BM_CONV_2D(1, 64, 56, 56, 1, 1, 1, VALID, 128, float);
+// BM_CONV_2D(1, 256, 28, 28, 1, 1, 1, VALID, 256, float);
+// BM_CONV_2D(1, 1024, 7, 7, 1, 1, 1, VALID, 1024, float);
+// BM_CONV_2D(1, 64, 32, 32, 3, 3, 1, VALID, 128, float);
+// BM_CONV_2D(1, 64, 33, 31, 3, 3, 1, VALID, 128, float);
+// BM_CONV_2D(1, 3, 512, 512, 3, 3, 1, VALID, 3, float);
+// BM_CONV_2D(1, 64, 32, 32, 3, 3, 1, SAME, 128, float);
+// BM_CONV_2D(1, 64, 33, 31, 3, 3, 1, SAME, 128, float);
+// BM_CONV_2D(1, 64, 32, 32, 3, 3, 2, VALID, 128, float);
+// BM_CONV_2D(1, 3, 512, 512, 3, 3, 2, VALID, 3, float);
+// BM_CONV_2D(1, 64, 33, 31, 3, 3, 2, VALID, 128, float);
+// BM_CONV_2D(1, 64, 32, 32, 3, 3, 2, SAME, 128, float);
+// BM_CONV_2D(1, 64, 33, 31, 3, 3, 2, SAME, 128, float);
+// BM_CONV_2D(1, 64, 32, 32, 5, 5, 1, VALID, 128, float);
+// BM_CONV_2D(1, 64, 32, 31, 5, 5, 1, VALID, 128, float);
+// BM_CONV_2D(1, 64, 32, 32, 5, 5, 1, SAME, 128, float);
+// BM_CONV_2D(1, 64, 32, 31, 5, 5, 1, SAME, 128, float);
 }  //  namespace mace

mace/ops/conv_2d_test.cc

@@ -10,7 +10,7 @@ using namespace mace;
 
 class Conv2dOpTest : public OpsTestBase {};
 
-template<DeviceType D>
+template <DeviceType D>
 void TestSimple3x3VALID() {
   OpsTestNet net;
   OpDefBuilder("Conv2D", "Conv2dTest")
@@ -42,10 +42,9 @@ void TestSimple3x3VALID() {
   auto expected = CreateTensor<float>({1, 1, 1, 1}, {18.1f});
 
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
-
 }
 
-template<DeviceType D>
+template <DeviceType D>
 void TestSimple3x3SAME() {
   OpsTestNet net;
   OpDefBuilder("Conv2D", "Conv2dTest")
@@ -86,7 +85,7 @@ TEST_F(Conv2dOpTest, NEONSimple) {
 }
 #endif
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 void TestNHWCSimple3x3VALID() {
   OpsTestNet net;
   // Add input data
@@ -100,9 +99,12 @@ void TestNHWCSimple3x3VALID() {
   net.AddInputFromArray<D, T>("Bias", {1}, {0.1f});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, T>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Filter", "FilterImage",
+                        kernels::BufferType::FILTER);
+    BufferToImage<D, T>(net, "Bias", "BiasImage",
+                        kernels::BufferType::ARGUMENT);
     OpDefBuilder("Conv2D", "Conv2dTest")
         .Input("InputImage")
         .Input("FilterImage")
@@ -117,7 +119,8 @@ void TestNHWCSimple3x3VALID() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, T>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, T>(net, "OutputImage", "Output",
+                        kernels::BufferType::IN_OUT);
 
   } else {
     OpDefBuilder("Conv2D", "Conv2dTest")
@@ -138,7 +141,7 @@ void TestNHWCSimple3x3VALID() {
   ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 0.01);
 }
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 void TestNHWCSimple3x3SAME() {
   OpsTestNet net;
 
@@ -153,9 +156,12 @@ void TestNHWCSimple3x3SAME() {
   net.AddInputFromArray<D, T>("Bias", {1}, {0.1f});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, T>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Filter", "FilterImage",
+                        kernels::BufferType::FILTER);
+    BufferToImage<D, T>(net, "Bias", "BiasImage",
+                        kernels::BufferType::ARGUMENT);
     OpDefBuilder("Conv2D", "Conv2dTest")
         .Input("InputImage")
         .Input("FilterImage")
@@ -170,7 +176,8 @@ void TestNHWCSimple3x3SAME() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, T>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, T>(net, "OutputImage", "Output",
+                        kernels::BufferType::IN_OUT);
 
   } else {
     OpDefBuilder("Conv2D", "Conv2dTest")
@@ -204,7 +211,7 @@ TEST_F(Conv2dOpTest, OPENCLSimple) {
   TestNHWCSimple3x3SAME<DeviceType::OPENCL, float>();
 }
 
-template<DeviceType D>
+template <DeviceType D>
 void TestSimple3x3WithoutBias() {
   OpsTestNet net;
   OpDefBuilder("Conv2D", "Conv2dTest")
@@ -234,14 +241,13 @@ void TestSimple3x3WithoutBias() {
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
 }
 
-
 #ifdef __ARM_NEON
 TEST_F(Conv2dOpTest, NEONWithouBias) {
   TestSimple3x3WithoutBias<DeviceType::NEON>();
 }
 #endif
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 void TestNHWCSimple3x3WithoutBias() {
   OpsTestNet net;
 
@@ -255,8 +261,10 @@ void TestNHWCSimple3x3WithoutBias() {
        1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Filter", "FilterImage",
+                        kernels::BufferType::FILTER);
 
     OpDefBuilder("Conv2D", "Conv2dTest")
         .Input("InputImage")
@@ -270,7 +278,8 @@ void TestNHWCSimple3x3WithoutBias() {
     // Run
     net.RunOp(D);
     // Transfer output
-    ImageToBuffer<D, T>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, T>(net, "OutputImage", "Output",
+                        kernels::BufferType::IN_OUT);
   } else {
     OpDefBuilder("Conv2D", "Conv2dTest")
         .Input("Input")
@@ -300,7 +309,7 @@ TEST_F(Conv2dOpTest, OPENCLWithoutBias) {
   TestNHWCSimple3x3WithoutBias<DeviceType::OPENCL, float>();
 }
 
-template<DeviceType D>
+template <DeviceType D>
 static void TestCombined3x3() {
   // Construct graph
   OpsTestNet net;
@@ -335,17 +344,13 @@ static void TestCombined3x3() {
                      4.2f, 6.2f, 4.2f, 6.2f, 9.2f, 6.2f, 4.2f, 6.2f, 4.2f});
 
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
-
 }
 
-
 #ifdef __ARM_NEON
-TEST_F(Conv2dOpTest, NEONCombined) {
-  TestCombined3x3<DeviceType::NEON>();
-}
+TEST_F(Conv2dOpTest, NEONCombined) { TestCombined3x3<DeviceType::NEON>(); }
 #endif
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 static void TestNHWCCombined3x3() {
   // Construct graph
   OpsTestNet net;
@@ -363,9 +368,12 @@ static void TestNHWCCombined3x3() {
   net.AddInputFromArray<D, T>("Bias", {2}, {0.1f, 0.2f});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, T>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Filter", "FilterImage",
+                        kernels::BufferType::FILTER);
+    BufferToImage<D, T>(net, "Bias", "BiasImage",
+                        kernels::BufferType::ARGUMENT);
 
     OpDefBuilder("Conv2D", "Conv2DTest")
         .Input("InputImage")
@@ -380,7 +388,8 @@ static void TestNHWCCombined3x3() {
     // Run
     net.RunOp(D);
 
-    ImageToBuffer<D, T>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, T>(net, "OutputImage", "Output",
+                        kernels::BufferType::IN_OUT);
   } else {
     OpDefBuilder("Conv2D", "Conv2DTest")
         .Input("Input")
@@ -394,16 +403,13 @@ static void TestNHWCCombined3x3() {
         .Finalize(net.NewOperatorDef());
     // Run
     net.RunOp(D);
-
   }
 
   // Check
   auto expected = CreateTensor<float>(
-      {1, 3, 3, 2}, {8.1f, 4.2f, 12.1f, 6.2f, 8.1f, 4.2f,
-                     12.1f, 6.2f, 18.1f, 9.2f, 12.1f, 6.2f,
-                     8.1f, 4.2f, 12.1f, 6.2f, 8.1f, 4.2f});
+      {1, 3, 3, 2}, {8.1f, 4.2f, 12.1f, 6.2f, 8.1f, 4.2f, 12.1f, 6.2f, 18.1f,
+                     9.2f, 12.1f, 6.2f, 8.1f, 4.2f, 12.1f, 6.2f, 8.1f, 4.2f});
   ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 0.01);
-
 }
 
 TEST_F(Conv2dOpTest, CPUStride2) {
@@ -414,7 +420,7 @@ TEST_F(Conv2dOpTest, OPENCLStride2) {
   TestNHWCCombined3x3<DeviceType::OPENCL, float>();
 }
 
-template<DeviceType D>
+template <DeviceType D>
 void TestConv1x1() {
   // Construct graph
   OpsTestNet net;
@@ -435,9 +441,12 @@ void TestConv1x1() {
   net.AddInputFromArray<D, float>("Bias", {2}, {0.1f, 0.2f});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, float>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, float>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, float>(net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT);
+    BufferToImage<D, float>(net, "Filter", "FilterImage",
+                            kernels::BufferType::FILTER);
+    BufferToImage<D, float>(net, "Bias", "BiasImage",
+                            kernels::BufferType::ARGUMENT);
 
     OpDefBuilder("Conv2D", "Conv2DTest")
         .Input("InputImage")
@@ -451,7 +460,8 @@ void TestConv1x1() {
     // Run
     net.RunOp(D);
 
-    ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT);
   } else {
     OpDefBuilder("Conv2D", "Conv2DTest")
         .Input("Input")
@@ -479,15 +489,11 @@ void TestConv1x1() {
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
 }
 
-TEST_F(Conv2dOpTest, CPUConv1x1) {
-  TestConv1x1<DeviceType::CPU>();
-}
+TEST_F(Conv2dOpTest, CPUConv1x1) { TestConv1x1<DeviceType::CPU>(); }
 
-TEST_F(Conv2dOpTest, OPENCLConv1x1) {
-  TestConv1x1<DeviceType::OPENCL>();
-}
+TEST_F(Conv2dOpTest, OPENCLConv1x1) { TestConv1x1<DeviceType::OPENCL>(); }
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 static void TestComplexConvNxNS12(const std::vector<index_t> &shape) {
   testing::internal::LogToStderr();
   auto func = [&](int kernel_h, int kernel_w, int stride_h, int stride_w,
@@ -526,9 +532,12 @@ static void TestComplexConvNxNS12(const std::vector<index_t> &shape) {
     expected.Copy(*net.GetOutput("Output"));
 
     // run on gpu
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, T>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Filter", "FilterImage",
+                        kernels::BufferType::FILTER);
+    BufferToImage<D, T>(net, "Bias", "BiasImage",
+                        kernels::BufferType::ARGUMENT);
 
     OpDefBuilder("Conv2D", "Conv2dTest")
         .Input("InputImage")
@@ -543,7 +552,8 @@ static void TestComplexConvNxNS12(const std::vector<index_t> &shape) {
     // Run on device
     net.RunOp(D);
 
-    ImageToBuffer<D, T>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, T>(net, "OutputImage", "OPENCLOutput",
+                        kernels::BufferType::IN_OUT);
     ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.001);
   };
 
@@ -592,15 +602,20 @@ static void TestHalfComplexConvNxNS12(const std::vector<index_t> &input_shape,
         .Finalize(net.NewOperatorDef());
 
     std::vector<float> float_input_data;
-    GenerateRandomRealTypeData({batch, height, width, input_channels}, float_input_data);
+    GenerateRandomRealTypeData({batch, height, width, input_channels},
+                               float_input_data);
     std::vector<float> float_filter_data;
-    GenerateRandomRealTypeData({kernel_h, kernel_w, input_channels, output_channels}, float_filter_data);
+    GenerateRandomRealTypeData(
+        {kernel_h, kernel_w, input_channels, output_channels},
+        float_filter_data);
     std::vector<float> float_bias_data;
     GenerateRandomRealTypeData({output_channels}, float_bias_data);
     // Add input data
-    net.AddInputFromArray<D, float>("Input", {batch, height, width, input_channels}, float_input_data);
     net.AddInputFromArray<D, float>(
-        "Filter", {kernel_h, kernel_w, input_channels, output_channels}, float_filter_data);
+        "Input", {batch, height, width, input_channels}, float_input_data);
+    net.AddInputFromArray<D, float>(
+        "Filter", {kernel_h, kernel_w, input_channels, output_channels},
+        float_filter_data);
     net.AddInputFromArray<D, float>("Bias", {output_channels}, float_bias_data);
 
     // run on cpu
@@ -610,9 +625,12 @@ static void TestHalfComplexConvNxNS12(const std::vector<index_t> &input_shape,
     expected.Copy(*net.GetOutput("Output"));
 
     // run on gpu
-    BufferToImage<D, half>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, half>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, half>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, half>(net, "Input", "InputImage",
+                           kernels::BufferType::IN_OUT);
+    BufferToImage<D, half>(net, "Filter", "FilterImage",
+                           kernels::BufferType::FILTER);
+    BufferToImage<D, half>(net, "Bias", "BiasImage",
+                           kernels::BufferType::ARGUMENT);
 
     OpDefBuilder("Conv2D", "Conv2dTest")
         .Input("InputImage")
@@ -627,7 +645,8 @@ static void TestHalfComplexConvNxNS12(const std::vector<index_t> &input_shape,
     // Run on device
     net.RunOp(D);
 
-    ImageToBuffer<D, float>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "OPENCLOutput",
+                            kernels::BufferType::IN_OUT);
 
     ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.5);
   };

mace/ops/core_test.cc

@@ -7,7 +7,6 @@
 namespace mace {
 
 TEST(CoreTest, INIT_MODE) {
-
   std::vector<OperatorDef> op_defs;
 
   Workspace ws;
@@ -18,10 +17,11 @@ TEST(CoreTest, INIT_MODE) {
       .Output("B2IOutput")
       .AddIntArg("buffer_type", kernels::BufferType::FILTER)
       .AddIntArg("mode", static_cast<int>(NetMode::INIT))
-      .Finalize(&op_defs[op_defs.size()-1]);
+      .Finalize(&op_defs[op_defs.size() - 1]);
 
   Tensor *input =
-      ws.CreateTensor("Input", GetDeviceAllocator(DeviceType::OPENCL), DataTypeToEnum<float>::v());
+      ws.CreateTensor("Input", GetDeviceAllocator(DeviceType::OPENCL),
+                      DataTypeToEnum<float>::v());
   input->Resize({1, 3, 3, 3});
   {
     Tensor::MappingGuard input_mapper(input);
@@ -34,23 +34,26 @@ TEST(CoreTest, INIT_MODE) {
       .Input("B2IOutput")
       .Output("Output")
       .AddIntArg("buffer_type", kernels::BufferType::FILTER)
-      .Finalize(&op_defs[op_defs.size()-1]);
+      .Finalize(&op_defs[op_defs.size() - 1]);
 
   NetDef net_def;
   for (auto &op_def : op_defs) {
     net_def.add_op()->CopyFrom(op_def);
   }
-  auto net = CreateNet(net_def, &ws, DeviceType::OPENCL, NetMode::INIT);
+  std::shared_ptr<OperatorRegistry> op_registry(new OperatorRegistry());
+  auto net =
+      CreateNet(op_registry, net_def, &ws, DeviceType::OPENCL, NetMode::INIT);
   net->Run();
 
   EXPECT_TRUE(ws.GetTensor("B2IOutput") != nullptr);
   EXPECT_TRUE(ws.GetTensor("Output") == nullptr);
 
-  net = CreateNet(net_def, &ws, DeviceType::OPENCL);
+  net = CreateNet(op_registry, net_def, &ws, DeviceType::OPENCL);
   net->Run();
   EXPECT_TRUE(ws.GetTensor("Output") != nullptr);
 
-  ExpectTensorNear<float>(*ws.GetTensor("Input"), *ws.GetTensor("Output"), 1e-5);
+  ExpectTensorNear<float>(*ws.GetTensor("Input"), *ws.GetTensor("Output"),
+                          1e-5);
 }
 
 }  //  namespace mace

mace/ops/depthwise_conv2d.cc

@@ -6,21 +6,26 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("DepthwiseConv2d")
+void Register_DepthwiseConv2d(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("DepthwiseConv2d")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     DepthwiseConv2dOp<DeviceType::CPU, float>);
 
 #if MACE_ENABLE_NEON
-REGISTER_NEON_OPERATOR(OpKeyBuilder("DepthwiseConv2d")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("DepthwiseConv2d")
+                                     .Device(DeviceType::NEON)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     DepthwiseConv2dOp<DeviceType::NEON, float>);
 #endif  // MACE_ENABLE_NEON
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("DepthwiseConv2d")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("DepthwiseConv2d")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     DepthwiseConv2dOp<DeviceType::OPENCL, float>);
+}
 
 }  // namespace mace

mace/ops/depthwise_conv2d_test.cc

@@ -41,12 +41,9 @@ void SimpleValidTest() {
       {196.1f, 252.1f, 216.2f, 280.2f, 272.3f, 344.3f, 296.4f, 376.4f});
 
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
-
 }
 
-TEST_F(DepthwiseConv2dOpTest, SimpleCPU) {
-  SimpleValidTest<DeviceType::CPU>();
-}
+TEST_F(DepthwiseConv2dOpTest, SimpleCPU) { SimpleValidTest<DeviceType::CPU>(); }
 
 template <DeviceType D>
 void TestNxNS12(const index_t height, const index_t width) {
@@ -72,8 +69,10 @@ void TestNxNS12(const index_t height, const index_t width) {
         .Finalize(net.NewOperatorDef());
 
     // Add input data
-    net.AddRandomInput<D, float>("Input", {batch, input_channels, height, width});
-    net.AddRandomInput<D, float>("Filter", {multiplier, input_channels, kernel_h, kernel_w});
+    net.AddRandomInput<D, float>("Input",
+                                 {batch, input_channels, height, width});
+    net.AddRandomInput<D, float>(
+        "Filter", {multiplier, input_channels, kernel_h, kernel_w});
     net.AddRandomInput<D, float>("Bias", {multiplier * input_channels});
     // Run on device
     net.RunOp(D);
@@ -93,7 +92,6 @@ void TestNxNS12(const index_t height, const index_t width) {
       func(kernel_size, kernel_size, stride, stride, SAME);
     }
   }
-
 }
 
 #if __ARM_NEON

mace/ops/depthwise_conv_2d_benchmark.cc

@@ -39,7 +39,7 @@ static void DepthwiseConv2d(int iters,
   net.AddRandomInput<D, float>("Input", {batch, channels, height, width});
   net.AddRandomInput<D, float>("Filter",
                                {output_channels, channels, kernel_h, kernel_w});
-  net.AddRandomInput<D, float>("Bias", {output_channels*channels});
+  net.AddRandomInput<D, float>("Bias", {output_channels * channels});
 
   // Warm-up
   for (int i = 0; i < 5; ++i) {
@@ -70,7 +70,6 @@ static void DepthwiseConv2d(int iters,
 
 #define BM_DEPTHWISE_CONV_2D(N, C, H, W, KH, KW, S, P, OC, TYPE)        \
   BM_DEPTHWISE_CONV_2D_MACRO(N, C, H, W, KH, KW, S, P, OC, TYPE, CPU);  \
-  BM_DEPTHWISE_CONV_2D_MACRO(N, C, H, W, KH, KW, S, P, OC, TYPE, NEON);\
   BM_DEPTHWISE_CONV_2D_MACRO(N, C, H, W, KH, KW, S, P, OC, TYPE, OPENCL);
 
 BM_DEPTHWISE_CONV_2D(1, 64, 32, 32, 3, 3, 1, VALID, 2, float);

mace/ops/fused_conv_2d.cc

@@ -6,25 +6,30 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("FusedConv2D")
+void Register_FusedConv2D(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("FusedConv2D")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     FusedConv2dOp<DeviceType::CPU, float>);
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("FusedConv2D")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("FusedConv2D")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     FusedConv2dOp<DeviceType::CPU, half>);
 
-
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("FusedConv2D")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("FusedConv2D")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     FusedConv2dOp<DeviceType::OPENCL, float>);
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("FusedConv2D")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("FusedConv2D")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     FusedConv2dOp<DeviceType::OPENCL, half>);
+}
 
 }  // namespace mace

mace/ops/fused_conv_2d_test.cc

@@ -9,7 +9,7 @@ using namespace mace;
 
 class FusedConv2dOpTest : public OpsTestBase {};
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 void TestNHWCSimple3x3VALID() {
   OpsTestNet net;
   // Add input data
@@ -23,9 +23,12 @@ void TestNHWCSimple3x3VALID() {
   net.AddInputFromArray<D, T>("Bias", {1}, {-0.1f});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, T>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Filter", "FilterImage",
+                        kernels::BufferType::FILTER);
+    BufferToImage<D, T>(net, "Bias", "BiasImage",
+                        kernels::BufferType::ARGUMENT);
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
         .Input("InputImage")
         .Input("FilterImage")
@@ -40,7 +43,8 @@ void TestNHWCSimple3x3VALID() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, T>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, T>(net, "OutputImage", "Output",
+                        kernels::BufferType::IN_OUT);
 
   } else {
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
@@ -61,7 +65,7 @@ void TestNHWCSimple3x3VALID() {
   ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 0.01);
 }
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 void TestNHWCSimple3x3SAME() {
   OpsTestNet net;
 
@@ -76,9 +80,12 @@ void TestNHWCSimple3x3SAME() {
   net.AddInputFromArray<D, T>("Bias", {1}, {-0.1f});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, T>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Filter", "FilterImage",
+                        kernels::BufferType::FILTER);
+    BufferToImage<D, T>(net, "Bias", "BiasImage",
+                        kernels::BufferType::ARGUMENT);
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
         .Input("InputImage")
         .Input("FilterImage")
@@ -93,7 +100,8 @@ void TestNHWCSimple3x3SAME() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, T>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, T>(net, "OutputImage", "Output",
+                        kernels::BufferType::IN_OUT);
 
   } else {
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
@@ -111,8 +119,7 @@ void TestNHWCSimple3x3SAME() {
   }
 
   auto expected = CreateTensor<float>(
-      {1, 3, 3, 1},
-      {0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f});
+      {1, 3, 3, 1}, {0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f});
 
   ExpectTensorNear<float, T>(*expected, *net.GetOutput("Output"), 0.01);
 }
@@ -127,7 +134,7 @@ TEST_F(FusedConv2dOpTest, OPENCLSimple) {
   TestNHWCSimple3x3SAME<DeviceType::OPENCL, float>();
 }
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 void TestNHWCSimple3x3WithoutBias() {
   OpsTestNet net;
 
@@ -141,8 +148,10 @@ void TestNHWCSimple3x3WithoutBias() {
        1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Filter", "FilterImage",
+                        kernels::BufferType::FILTER);
 
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
         .Input("InputImage")
@@ -156,7 +165,8 @@ void TestNHWCSimple3x3WithoutBias() {
     // Run
     net.RunOp(D);
     // Transfer output
-    ImageToBuffer<D, T>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, T>(net, "OutputImage", "Output",
+                        kernels::BufferType::IN_OUT);
   } else {
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
         .Input("Input")
@@ -186,7 +196,7 @@ TEST_F(FusedConv2dOpTest, OPENCLWithoutBias) {
   TestNHWCSimple3x3WithoutBias<DeviceType::OPENCL, float>();
 }
 
-template<DeviceType D>
+template <DeviceType D>
 void TestConv1x1() {
   // Construct graph
   OpsTestNet net;
@@ -207,9 +217,12 @@ void TestConv1x1() {
   net.AddInputFromArray<D, float>("Bias", {2}, {0.1f, 0.2f});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, float>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, float>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, float>(net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT);
+    BufferToImage<D, float>(net, "Filter", "FilterImage",
+                            kernels::BufferType::FILTER);
+    BufferToImage<D, float>(net, "Bias", "BiasImage",
+                            kernels::BufferType::ARGUMENT);
 
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
         .Input("InputImage")
@@ -223,7 +236,8 @@ void TestConv1x1() {
     // Run
     net.RunOp(D);
 
-    ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT);
   } else {
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
         .Input("Input")
@@ -251,15 +265,11 @@ void TestConv1x1() {
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
 }
 
-TEST_F(FusedConv2dOpTest, CPUConv1x1) {
-  TestConv1x1<DeviceType::CPU>();
-}
+TEST_F(FusedConv2dOpTest, CPUConv1x1) { TestConv1x1<DeviceType::CPU>(); }
 
-TEST_F(FusedConv2dOpTest, OPENCLConv1x1) {
-  TestConv1x1<DeviceType::OPENCL>();
-}
+TEST_F(FusedConv2dOpTest, OPENCLConv1x1) { TestConv1x1<DeviceType::OPENCL>(); }
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 static void TestComplexConvNxNS12(const std::vector<index_t> &shape) {
   testing::internal::LogToStderr();
   auto func = [&](int kernel_h, int kernel_w, int stride_h, int stride_w,
@@ -298,9 +308,12 @@ static void TestComplexConvNxNS12(const std::vector<index_t> &shape) {
     expected.Copy(*net.GetOutput("Output"));
 
     // run on gpu
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, T>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Filter", "FilterImage",
+                        kernels::BufferType::FILTER);
+    BufferToImage<D, T>(net, "Bias", "BiasImage",
+                        kernels::BufferType::ARGUMENT);
 
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
         .Input("InputImage")
@@ -315,7 +328,8 @@ static void TestComplexConvNxNS12(const std::vector<index_t> &shape) {
     // Run on device
     net.RunOp(D);
 
-    ImageToBuffer<D, T>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, T>(net, "OutputImage", "OPENCLOutput",
+                        kernels::BufferType::IN_OUT);
     ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.001);
   };
 
@@ -331,7 +345,7 @@ TEST_F(FusedConv2dOpTest, OPENCLUnalignedConvNxNS12) {
   TestComplexConvNxNS12<DeviceType::OPENCL, float>({107, 113, 5, 7});
 }
 
-template<DeviceType D>
+template <DeviceType D>
 static void TestHalfComplexConvNxNS12(const std::vector<index_t> &shape) {
   testing::internal::LogToStderr();
   auto func = [&](int kernel_h, int kernel_w, int stride_h, int stride_w,
@@ -357,15 +371,20 @@ static void TestHalfComplexConvNxNS12(const std::vector<index_t> &shape) {
         .Finalize(net.NewOperatorDef());
 
     std::vector<float> float_input_data;
-    GenerateRandomRealTypeData({batch, height, width, input_channels}, float_input_data);
+    GenerateRandomRealTypeData({batch, height, width, input_channels},
+                               float_input_data);
     std::vector<float> float_filter_data;
-    GenerateRandomRealTypeData({kernel_h, kernel_w, input_channels, output_channels}, float_filter_data);
+    GenerateRandomRealTypeData(
+        {kernel_h, kernel_w, input_channels, output_channels},
+        float_filter_data);
     std::vector<float> float_bias_data;
     GenerateRandomRealTypeData({output_channels}, float_bias_data);
     // Add input data
-    net.AddInputFromArray<D, float>("Input", {batch, height, width, input_channels}, float_input_data);
     net.AddInputFromArray<D, float>(
-        "Filter", {kernel_h, kernel_w, input_channels, output_channels}, float_filter_data);
+        "Input", {batch, height, width, input_channels}, float_input_data);
+    net.AddInputFromArray<D, float>(
+        "Filter", {kernel_h, kernel_w, input_channels, output_channels},
+        float_filter_data);
     net.AddInputFromArray<D, float>("Bias", {output_channels}, float_bias_data);
 
     // run on cpu
@@ -375,9 +394,12 @@ static void TestHalfComplexConvNxNS12(const std::vector<index_t> &shape) {
     expected.Copy(*net.GetOutput("Output"));
 
     // run on gpu
-    BufferToImage<D, half>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, half>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, half>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, half>(net, "Input", "InputImage",
+                           kernels::BufferType::IN_OUT);
+    BufferToImage<D, half>(net, "Filter", "FilterImage",
+                           kernels::BufferType::FILTER);
+    BufferToImage<D, half>(net, "Bias", "BiasImage",
+                           kernels::BufferType::ARGUMENT);
 
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
         .Input("InputImage")
@@ -392,7 +414,8 @@ static void TestHalfComplexConvNxNS12(const std::vector<index_t> &shape) {
     // Run on device
     net.RunOp(D);
 
-    ImageToBuffer<D, float>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "OPENCLOutput",
+                            kernels::BufferType::IN_OUT);
 
     ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.2);
   };
@@ -408,7 +431,7 @@ TEST_F(FusedConv2dOpTest, OPENCLHalfAlignedConvNxNS12) {
   TestHalfComplexConvNxNS12<DeviceType::OPENCL>({32, 32, 32, 64});
 }
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 static void TestGeneralConvNxNS12(const std::vector<index_t> &image_shape,
                                   const std::vector<index_t> &filter_shape) {
   testing::internal::LogToStderr();
@@ -449,9 +472,12 @@ static void TestGeneralConvNxNS12(const std::vector<index_t> &image_shape,
     expected.Copy(*net.GetOutput("Output"));
 
     // run on gpu
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
-    BufferToImage<D, T>(net, "Filter", "FilterImage", kernels::BufferType::FILTER);
-    BufferToImage<D, T>(net, "Bias", "BiasImage", kernels::BufferType::ARGUMENT);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Filter", "FilterImage",
+                        kernels::BufferType::FILTER);
+    BufferToImage<D, T>(net, "Bias", "BiasImage",
+                        kernels::BufferType::ARGUMENT);
 
     OpDefBuilder("FusedConv2D", "FusedConv2dTest")
         .Input("InputImage")
@@ -466,7 +492,8 @@ static void TestGeneralConvNxNS12(const std::vector<index_t> &image_shape,
     // Run on device
     net.RunOp(D);
 
-    ImageToBuffer<D, T>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, T>(net, "OutputImage", "OPENCLOutput",
+                        kernels::BufferType::IN_OUT);
     ExpectTensorNear<float>(expected, *net.GetOutput("OPENCLOutput"), 0.001);
   };
 
@@ -477,13 +504,11 @@ static void TestGeneralConvNxNS12(const std::vector<index_t> &image_shape,
 }
 
 TEST_F(FusedConv2dOpTest, OPENCL7X7ConvNxNS12) {
-  TestGeneralConvNxNS12<DeviceType::OPENCL, float>({32, 32},
-                                                   {7, 7, 3, 64});
+  TestGeneralConvNxNS12<DeviceType::OPENCL, float>({32, 32}, {7, 7, 3, 64});
 }
 
 TEST_F(FusedConv2dOpTest, OPENCL15X1ConvNxNS12) {
-  TestGeneralConvNxNS12<DeviceType::OPENCL, float>({40, 40},
-                                                   {15, 1, 32, 64});
+  TestGeneralConvNxNS12<DeviceType::OPENCL, float>({40, 40}, {15, 1, 32, 64});
 }
 
 template<DeviceType D, typename T>

mace/ops/global_avg_pooling.cc

@@ -6,16 +6,20 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("GlobalAvgPooling")
+void Register_GlobalAvgPooling(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("GlobalAvgPooling")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     GlobalAvgPoolingOp<DeviceType::CPU, float>);
 
 #if MACE_ENABLE_NEON
-REGISTER_NEON_OPERATOR(OpKeyBuilder("GlobalAvgPooling")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("GlobalAvgPooling")
+                                     .Device(DeviceType::NEON)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     GlobalAvgPoolingOp<DeviceType::NEON, float>);
 #endif  // MACE_ENABLE_NEON
+}
 
 }  //  namespace mace

mace/ops/global_avg_pooling_benchmark.cc

@@ -22,7 +22,8 @@ static void GlobalAvgPooling(
       .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<DeviceType::CPU, float>("Input", {batch, channels, height, width});
+  net.AddRandomInput<DeviceType::CPU, float>("Input",
+                                             {batch, channels, height, width});
 
   // Warm-up
   for (int i = 0; i < 5; ++i) {

mace/ops/image_to_buffer.cc

@@ -6,14 +6,18 @@
 
 namespace mace {
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("ImageToBuffer")
+void Register_ImageToBuffer(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("ImageToBuffer")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     ImageToBufferOp<DeviceType::OPENCL, float>);
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("ImageToBuffer")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("ImageToBuffer")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     ImageToBufferOp<DeviceType::OPENCL, half>);
+}
 
 }  //  namespace mace

mace/ops/ops_test_util.h

@@ -10,9 +10,9 @@
 #include "gtest/gtest.h"
 #include "mace/core/common.h"
 #include "mace/core/net.h"
+#include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/core/tensor.h"
 #include "mace/core/workspace.h"
-#include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/kernels/opencl/helper.h"
 #include "mace/utils/utils.h"
 
@@ -56,7 +56,8 @@ class OpDefBuilder {
     return *this;
   }
 
-  OpDefBuilder AddIntsArg(const std::string &name, const std::vector<int> &values) {
+  OpDefBuilder AddIntsArg(const std::string &name,
+                          const std::vector<int> &values) {
     auto arg = op_def_.add_arg();
     arg->set_name(name);
     for (auto value : values) {
@@ -65,7 +66,8 @@ class OpDefBuilder {
     return *this;
   }
 
-  OpDefBuilder AddFloatsArg(const std::string &name, const std::vector<float> &values) {
+  OpDefBuilder AddFloatsArg(const std::string &name,
+                            const std::vector<float> &values) {
     auto arg = op_def_.add_arg();
     arg->set_name(name);
     for (auto value : values) {
@@ -94,7 +96,7 @@ class OpDefBuilder {
 
 class OpsTestNet {
  public:
-  OpsTestNet() {}
+  OpsTestNet() : op_registry_(new OperatorRegistry()) {};
 
   template <DeviceType D, typename T>
   void AddInputFromArray(const std::string &name,
@@ -135,9 +137,10 @@ class OpsTestNet {
     std::mt19937 gen(rd());
     std::normal_distribution<float> nd(0, 1);
     if (DataTypeToEnum<T>::value == DT_HALF) {
-      std::generate(input_data, input_data + input->size(),
-                    [&gen, &nd, positive] {
-                      return half_float::half_cast<half>(positive ? std::abs(nd(gen)) : nd(gen));
+      std::generate(
+          input_data, input_data + input->size(), [&gen, &nd, positive] {
+            return half_float::half_cast<half>(positive ? std::abs(nd(gen))
+                                                        : nd(gen));
           });
     } else {
       std::generate(input_data, input_data + input->size(),
@@ -160,7 +163,7 @@ class OpsTestNet {
     for (auto &op_def_ : op_defs_) {
       net_def.add_op()->CopyFrom(op_def_);
     }
-    net_ = CreateNet(net_def, &ws_, device);
+    net_ = CreateNet(op_registry_, net_def, &ws_, device);
     device_ = device;
     return net_->Run();
   }
@@ -182,6 +185,7 @@ class OpsTestNet {
   }
 
  public:
+  std::shared_ptr<OperatorRegistry> op_registry_;
   Workspace ws_;
   std::vector<OperatorDef> op_defs_;
   std::unique_ptr<NetBase> net_;
@@ -211,7 +215,8 @@ void GenerateRandomRealTypeData(const std::vector<index_t> &shape,
   res.resize(size);
 
   if (DataTypeToEnum<T>::value == DT_HALF) {
-    std::generate(res.begin(), res.end(), [&gen, &nd] { return half_float::half_cast<half>(nd(gen)); });
+    std::generate(res.begin(), res.end(),
+                  [&gen, &nd] { return half_float::half_cast<half>(nd(gen)); });
   } else {
     std::generate(res.begin(), res.end(), [&gen, &nd] { return nd(gen); });
   }
@@ -236,7 +241,8 @@ void GenerateRandomIntTypeData(const std::vector<index_t> &shape,
 template <typename T>
 unique_ptr<Tensor> CreateTensor(const std::vector<index_t> &shape,
                                 const std::vector<T> &data) {
-  unique_ptr<Tensor> res(new Tensor(GetDeviceAllocator(DeviceType::CPU), DataTypeToEnum<T>::v()));
+  unique_ptr<Tensor> res(
+      new Tensor(GetDeviceAllocator(DeviceType::CPU), DataTypeToEnum<T>::v()));
   res->Resize(shape);
   T *input_data = res->mutable_data<T>();
   memcpy(input_data, data.data(), data.size() * sizeof(T));
@@ -268,9 +274,9 @@ inline std::string ShapeToString(const Tensor &x) {
 
 template <typename T>
 struct is_floating_point_type {
-  static const bool value =
-      std::is_same<T, float>::value || std::is_same<T, double>::value
-          || std::is_same<T, half>::value;
+  static const bool value = std::is_same<T, float>::value ||
+                            std::is_same<T, double>::value ||
+                            std::is_same<T, half>::value;
 };
 
 template <typename T>
@@ -293,7 +299,9 @@ inline void AssertSameDims(const Tensor &x, const Tensor &y) {
                                 << "y.shape [ " << ShapeToString(y) << "]";
 }
 
-template <typename EXP_TYPE, typename RES_TYPE, bool is_fp = is_floating_point_type<EXP_TYPE>::value>
+template <typename EXP_TYPE,
+          typename RES_TYPE,
+          bool is_fp = is_floating_point_type<EXP_TYPE>::value>
 struct Expector;
 
 // Partial specialization for float and double.
@@ -343,7 +351,6 @@ struct Expector<EXP_TYPE, RES_TYPE, true> {
       }
     }
   }
-
 };
 
 template <typename T>
@@ -355,8 +362,8 @@ void ExpectTensorNear(const Tensor &x, const Tensor &y, const double abs_err) {
 
 template <typename EXP_TYPE, typename RES_TYPE>
 void ExpectTensorNear(const Tensor &x, const Tensor &y, const double abs_err) {
-  static_assert(is_floating_point_type<EXP_TYPE>::value
-                    && is_floating_point_type<RES_TYPE>::value,
+  static_assert(is_floating_point_type<EXP_TYPE>::value &&
+                    is_floating_point_type<RES_TYPE>::value,
                 "T is not a floating point type");
   Expector<EXP_TYPE, RES_TYPE>::Near(x, y, abs_err);
 }

mace/ops/pooling.cc

@@ -6,29 +6,36 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("Pooling")
+void Register_Pooling(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Pooling")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     PoolingOp<DeviceType::CPU, float>);
-REGISTER_CPU_OPERATOR(OpKeyBuilder("Pooling")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Pooling")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     PoolingOp<DeviceType::CPU, half>);
 
 #if MACE_ENABLE_NEON
-REGISTER_NEON_OPERATOR(OpKeyBuilder("Pooling")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Pooling")
+                                     .Device(DeviceType::NEON)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     PoolingOp<DeviceType::NEON, float>);
 #endif  // MACE_ENABLE_NEON
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Pooling")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Pooling")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     PoolingOp<DeviceType::OPENCL, float>);
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Pooling")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Pooling")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     PoolingOp<DeviceType::OPENCL, half>);
+}
 
 }  //  namespace mace

mace/ops/pooling_benchmark.cc

@@ -35,7 +35,8 @@ static void Pooling(int iters,
       .Finalize(net.NewOperatorDef());
 
   // Add input data
-  net.AddRandomInput<DeviceType::CPU, float>("Input", {batch, channels, height, width});
+  net.AddRandomInput<DeviceType::CPU, float>("Input",
+                                             {batch, channels, height, width});
 
   // Warm-up
   for (int i = 0; i < 5; ++i) {

mace/ops/pooling_test.cc

@@ -42,7 +42,6 @@ TEST_F(PoolingOpTest, MAX_VALID) {
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
 }
 
-
 TEST_F(PoolingOpTest, MAX_SAME) {
   // Construct graph
   OpsTestNet net;
@@ -122,7 +121,7 @@ TEST_F(PoolingOpTest, MAX_k2x2s2x2) {
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
 }
 
-template<DeviceType D>
+template <DeviceType D>
 static void SimpleMaxPooling3S2() {
   // Construct graph
   OpsTestNet net;
@@ -134,7 +133,8 @@ static void SimpleMaxPooling3S2() {
        14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+    BufferToImage<D, float>(net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT);
     OpDefBuilder("Pooling", "PoolingTest")
         .Input("InputImage")
         .Output("OutputImage")
@@ -145,7 +145,8 @@ static void SimpleMaxPooling3S2() {
         .AddIntsArg("dilations", {1, 1})
         .Finalize(net.NewOperatorDef());
     net.RunOp(D);
-    ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT);
   } else {
     // Run
     OpDefBuilder("Pooling", "PoolingTest")
@@ -166,15 +167,13 @@ static void SimpleMaxPooling3S2() {
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
 }
 
-TEST_F(PoolingOpTest, CPUSimpleMaxPooling3S2) {
-  SimpleMaxPooling3S2<CPU>();
-}
+TEST_F(PoolingOpTest, CPUSimpleMaxPooling3S2) { SimpleMaxPooling3S2<CPU>(); }
 
 TEST_F(PoolingOpTest, OPENCLSimpleMaxPooling3S2) {
   SimpleMaxPooling3S2<OPENCL>();
 }
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 static void MaxPooling3S2(const std::vector<index_t> &input_shape,
                           const std::vector<int> strides,
                           Padding padding) {
@@ -211,13 +210,14 @@ static void MaxPooling3S2(const std::vector<index_t> &input_shape,
       .AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
       .Finalize(net.NewOperatorDef());
   net.RunOp(D);
-  ImageToBuffer<D, T>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+  ImageToBuffer<D, T>(net, "OutputImage", "OPENCLOutput",
+                      kernels::BufferType::IN_OUT);
 
   ExpectTensorNear<T>(expected, *net.GetOutput("OPENCLOutput"), 0.001);
 }
 
 // TODO(chenghui) : there is a bug.
-//TEST_F(PoolingOpTest, NEONAlignedMaxPooling3S2) {
+// TEST_F(PoolingOpTest, NEONAlignedMaxPooling3S2) {
 //  AlignedMaxPooling3S2<NEON>(Padding::VALID);
 //  AlignedMaxPooling3S2<NEON>(Padding::SAME);
 //}
@@ -272,7 +272,7 @@ TEST_F(PoolingOpTest, AVG_VALID) {
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 0.001);
 }
 
-template<DeviceType D>
+template <DeviceType D>
 static void SimpleAvgPoolingTest() {
   // Construct graph
   OpsTestNet net;
@@ -282,7 +282,8 @@ static void SimpleAvgPoolingTest() {
       "Input", {1, 2, 8, 1},
       {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15});
 
-  BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+  BufferToImage<D, float>(net, "Input", "InputImage",
+                          kernels::BufferType::IN_OUT);
   OpDefBuilder("Pooling", "PoolingTest")
       .Input("InputImage")
       .Output("OutputImage")
@@ -294,7 +295,8 @@ static void SimpleAvgPoolingTest() {
       .Finalize(net.NewOperatorDef());
   // Run
   net.RunOp(D);
-  ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+  ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                          kernels::BufferType::IN_OUT);
 
   // Check
   auto expected = CreateTensor<float>({1, 1, 4, 1}, {4.5, 6.5, 8.5, 10.5});
@@ -306,7 +308,7 @@ TEST_F(PoolingOpTest, OPENCLSimpleAvgPooling) {
   SimpleAvgPoolingTest<OPENCL>();
 }
 
-template<DeviceType D, typename T>
+template <DeviceType D, typename T>
 static void AvgPoolingTest(const std::vector<index_t> &shape,
                            const std::vector<int> &kernels,
                            const std::vector<int> &strides,
@@ -343,38 +345,49 @@ static void AvgPoolingTest(const std::vector<index_t> &shape,
       .AddIntArg("T", static_cast<int>(DataTypeToEnum<T>::value))
       .Finalize(net.NewOperatorDef());
   net.RunOp(D);
-  ImageToBuffer<D, T>(net, "OutputImage", "OPENCLOutput", kernels::BufferType::IN_OUT);
+  ImageToBuffer<D, T>(net, "OutputImage", "OPENCLOutput",
+                      kernels::BufferType::IN_OUT);
 
   ExpectTensorNear<float, T>(expected, *net.GetOutput("OPENCLOutput"), 0.01);
 }
 
 TEST_F(PoolingOpTest, OPENCLAlignedAvgPooling) {
-  AvgPoolingTest<OPENCL, float>({3, 15, 15, 128}, {4, 4}, {4, 4}, Padding::VALID);
-  AvgPoolingTest<OPENCL, float>({3, 15, 15, 128}, {4, 4}, {4, 4}, Padding::SAME);
+  AvgPoolingTest<OPENCL, float>({3, 15, 15, 128}, {4, 4}, {4, 4},
+                                Padding::VALID);
+  AvgPoolingTest<OPENCL, float>({3, 15, 15, 128}, {4, 4}, {4, 4},
+                                Padding::SAME);
 }
 
 TEST_F(PoolingOpTest, OPENCLHalfAlignedAvgPooling) {
-  AvgPoolingTest<OPENCL, half>({3, 15, 15, 128}, {4, 4}, {4, 4}, Padding::VALID);
+  AvgPoolingTest<OPENCL, half>({3, 15, 15, 128}, {4, 4}, {4, 4},
+                               Padding::VALID);
   AvgPoolingTest<OPENCL, half>({3, 15, 15, 128}, {4, 4}, {4, 4}, Padding::SAME);
 }
 
 TEST_F(PoolingOpTest, OPENCLAlignedLargeKernelAvgPooling) {
-  AvgPoolingTest<OPENCL, float>({3, 64, 64, 128}, {16, 16}, {16, 16}, Padding::VALID);
-  AvgPoolingTest<OPENCL, float>({3, 64, 64, 128}, {16, 16}, {16, 16}, Padding::SAME);
+  AvgPoolingTest<OPENCL, float>({3, 64, 64, 128}, {16, 16}, {16, 16},
+                                Padding::VALID);
+  AvgPoolingTest<OPENCL, float>({3, 64, 64, 128}, {16, 16}, {16, 16},
+                                Padding::SAME);
 }
 
 TEST_F(PoolingOpTest, OPENCLHalfAlignedLargeKernelAvgPooling) {
-  AvgPoolingTest<OPENCL, half>({3, 64, 64, 128}, {16, 16}, {16, 16}, Padding::VALID);
-  AvgPoolingTest<OPENCL, half>({3, 64, 64, 128}, {16, 16}, {16, 16}, Padding::SAME);
+  AvgPoolingTest<OPENCL, half>({3, 64, 64, 128}, {16, 16}, {16, 16},
+                               Padding::VALID);
+  AvgPoolingTest<OPENCL, half>({3, 64, 64, 128}, {16, 16}, {16, 16},
+                               Padding::SAME);
 }
 
 TEST_F(PoolingOpTest, OPENCLUnAlignedAvgPooling) {
-  AvgPoolingTest<OPENCL, float>({3, 31, 37, 128}, {2, 2}, {2, 2}, Padding::VALID);
-  AvgPoolingTest<OPENCL, float>({3, 31, 37, 128}, {2, 2}, {2, 2}, Padding::SAME);
+  AvgPoolingTest<OPENCL, float>({3, 31, 37, 128}, {2, 2}, {2, 2},
+                                Padding::VALID);
+  AvgPoolingTest<OPENCL, float>({3, 31, 37, 128}, {2, 2}, {2, 2},
+                                Padding::SAME);
 }
 
 TEST_F(PoolingOpTest, OPENCLUnAlignedLargeKernelAvgPooling) {
-  AvgPoolingTest<OPENCL, float>({3, 31, 37, 128}, {8, 8}, {8, 8}, Padding::VALID);
-  AvgPoolingTest<OPENCL, float>({3, 31, 37, 128}, {8, 8}, {8, 8}, Padding::SAME);
+  AvgPoolingTest<OPENCL, float>({3, 31, 37, 128}, {8, 8}, {8, 8},
+                                Padding::VALID);
+  AvgPoolingTest<OPENCL, float>({3, 31, 37, 128}, {8, 8}, {8, 8},
+                                Padding::SAME);
 }
-

mace/ops/relu.cc

@@ -6,26 +6,32 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("Relu")
+void Register_Relu(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Relu")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     ReluOp<DeviceType::CPU, float>);
 
 #if MACE_ENABLE_NEON
-REGISTER_NEON_OPERATOR(OpKeyBuilder("Relu")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Relu")
+                                     .Device(DeviceType::NEON)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     ReluOp<DeviceType::NEON, float>);
 #endif  // MACE_ENABLE_NEON
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Relu")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Relu")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     ReluOp<DeviceType::OPENCL, float>);
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Relu")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("Relu")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     ReluOp<DeviceType::OPENCL, half>);
+}
 
 }  //  namespace mace

mace/ops/relu_benchmark.cc

@@ -19,7 +19,8 @@ static void ReluBenchmark(
   net.AddRandomInput<D, float>("Input", {batch, height, width, channels});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+    BufferToImage<D, float>(net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT);
 
     OpDefBuilder("Relu", "ReluBM")
         .Input("InputImage")
@@ -56,7 +57,7 @@ static void ReluBenchmark(
 
 #define BM_RELU(N, C, H, W, TYPE)        \
   BM_RELU_MACRO(N, C, H, W, TYPE, CPU);  \
-  BM_RELU_MACRO(N, C, H, W, TYPE, NEON);\
+  BM_RELU_MACRO(N, C, H, W, TYPE, NEON); \
   BM_RELU_MACRO(N, C, H, W, TYPE, OPENCL);
 
 BM_RELU(1, 1, 512, 512, float);

mace/ops/relu_test.cc

@@ -14,13 +14,13 @@ void TestSimple() {
   OpsTestNet net;
 
   // Add input data
-  net.AddInputFromArray<D, float>("Input",
-                                  {2, 2, 2, 2},
-                                  {-7, 7, -6, 6, -5, 5, -4, 4,
-                                   -3, 3, -2, 2, -1, 1, 0, 0});
+  net.AddInputFromArray<D, float>(
+      "Input", {2, 2, 2, 2},
+      {-7, 7, -6, 6, -5, 5, -4, 4, -3, 3, -2, 2, -1, 1, 0, 0});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+    BufferToImage<D, float>(net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT);
 
     OpDefBuilder("Relu", "ReluTest")
         .Input("InputImage")
@@ -31,7 +31,8 @@ void TestSimple() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT);
   } else {
     OpDefBuilder("Relu", "ReluTest")
         .Input("Input")
@@ -42,38 +43,30 @@ void TestSimple() {
     net.RunOp(D);
   }
 
-  auto expected = CreateTensor<float>({2, 2, 2, 2},
-                                      {0, 7, 0, 6, 0, 5, 0, 4,
-                                       0, 3, 0, 2, 0, 1, 0, 0});
+  auto expected = CreateTensor<float>(
+      {2, 2, 2, 2}, {0, 7, 0, 6, 0, 5, 0, 4, 0, 3, 0, 2, 0, 1, 0, 0});
 
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
-TEST_F(ReluOpTest, CPUSimple) {
-  TestSimple<DeviceType::CPU>();
-}
+TEST_F(ReluOpTest, CPUSimple) { TestSimple<DeviceType::CPU>(); }
 
 #if __ARM_NEON
-TEST_F(ReluOpTest, NEONSimple) {
-  TestSimple<DeviceType::NEON>();
-}
+TEST_F(ReluOpTest, NEONSimple) { TestSimple<DeviceType::NEON>(); }
 #endif
 
-TEST_F(ReluOpTest, OPENCLSimple) {
-  TestSimple<DeviceType::OPENCL>();
-}
+TEST_F(ReluOpTest, OPENCLSimple) { TestSimple<DeviceType::OPENCL>(); }
 
 template <DeviceType D>
 void TestUnalignedSimple() {
   OpsTestNet net;
 
   // Add input data
-  net.AddInputFromArray<D, float>("Input",
-                                  {1, 3, 2, 1},
-                                  {-7, 7, -6, 6, -5, 5});
+  net.AddInputFromArray<D, float>("Input", {1, 3, 2, 1}, {-7, 7, -6, 6, -5, 5});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+    BufferToImage<D, float>(net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT);
 
     OpDefBuilder("Relu", "ReluTest")
         .Input("InputImage")
@@ -84,7 +77,8 @@ void TestUnalignedSimple() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT);
   } else {
     OpDefBuilder("Relu", "ReluTest")
         .Input("Input")
@@ -95,8 +89,7 @@ void TestUnalignedSimple() {
     net.RunOp(D);
   }
 
-  auto expected = CreateTensor<float>({1, 3, 2, 1},
-                                      {0, 7, 0, 6, 0, 5});
+  auto expected = CreateTensor<float>({1, 3, 2, 1}, {0, 7, 0, 6, 0, 5});
 
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
@@ -120,13 +113,13 @@ void TestSimpleReluX() {
   OpsTestNet net;
 
   // Add input data
-  net.AddInputFromArray<D, float>("Input",
-                                  {2, 2, 2, 2},
-                                  {-7, 7, -6, 6, -5, 5, -4, 4,
-                                   -3, 3, -2, 2, -1, 1, 0, 0});
+  net.AddInputFromArray<D, float>(
+      "Input", {2, 2, 2, 2},
+      {-7, 7, -6, 6, -5, 5, -4, 4, -3, 3, -2, 2, -1, 1, 0, 0});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+    BufferToImage<D, float>(net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT);
 
     OpDefBuilder("Relu", "ReluTest")
         .Input("InputImage")
@@ -138,7 +131,8 @@ void TestSimpleReluX() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT);
   } else {
     OpDefBuilder("Relu", "ReluTest")
         .Input("Input")
@@ -150,38 +144,31 @@ void TestSimpleReluX() {
     net.RunOp(D);
   }
 
-  auto expected = CreateTensor<float>({2, 2, 2, 2},
-                                      {0, 6, 0, 6, 0, 5, 0, 4,
-                                       0, 3, 0, 2, 0, 1, 0, 0});
+  auto expected = CreateTensor<float>(
+      {2, 2, 2, 2}, {0, 6, 0, 6, 0, 5, 0, 4, 0, 3, 0, 2, 0, 1, 0, 0});
 
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }
 
-TEST_F(ReluOpTest, CPUSimpleReluX) {
-  TestSimpleReluX<DeviceType::CPU>();
-}
+TEST_F(ReluOpTest, CPUSimpleReluX) { TestSimpleReluX<DeviceType::CPU>(); }
 
 #if __ARM_NEON
-TEST_F(ReluOpTest, NEONSimpleReluX) {
-  TestSimpleReluX<DeviceType::NEON>();
-}
+TEST_F(ReluOpTest, NEONSimpleReluX) { TestSimpleReluX<DeviceType::NEON>(); }
 #endif
 
-TEST_F(ReluOpTest, OPENCLSimpleReluX) {
-  TestSimpleReluX<DeviceType::OPENCL>();
-}
+TEST_F(ReluOpTest, OPENCLSimpleReluX) { TestSimpleReluX<DeviceType::OPENCL>(); }
 
 template <DeviceType D>
 void TestUnalignedSimpleReluX() {
   OpsTestNet net;
 
   // Add input data
-  net.AddInputFromArray<D, float>("Input",
-                                  {1, 1, 7, 1},
+  net.AddInputFromArray<D, float>("Input", {1, 1, 7, 1},
                                   {-7, 7, -6, 6, -5, 5, -4});
 
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+    BufferToImage<D, float>(net, "Input", "InputImage",
+                            kernels::BufferType::IN_OUT);
 
     OpDefBuilder("Relu", "ReluTest")
         .Input("InputImage")
@@ -193,7 +180,8 @@ void TestUnalignedSimpleReluX() {
     net.RunOp(D);
 
     // Transfer output
-    ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+    ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                            kernels::BufferType::IN_OUT);
   } else {
     OpDefBuilder("Relu", "ReluTest")
         .Input("Input")
@@ -205,8 +193,7 @@ void TestUnalignedSimpleReluX() {
     net.RunOp(D);
   }
 
-  auto expected = CreateTensor<float>({1, 1, 7, 1},
-                                      {0, 6, 0, 6, 0, 5, 0});
+  auto expected = CreateTensor<float>({1, 1, 7, 1}, {0, 6, 0, 6, 0, 5, 0});
 
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-5);
 }

mace/ops/resize_bilinear.cc

@@ -6,26 +6,32 @@
 
 namespace mace {
 
-REGISTER_CPU_OPERATOR(OpKeyBuilder("ResizeBilinear")
+void Register_ResizeBilinear(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("ResizeBilinear")
+                                     .Device(DeviceType::CPU)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     ResizeBilinearOp<DeviceType::CPU, float>);
 
 #if MACE_ENABLE_NEON
-REGISTER_NEON_OPERATOR(OpKeyBuilder("ResizeBilinear")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("ResizeBilinear")
+                                     .Device(DeviceType::NEON)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     ResizeBilinearOp<DeviceType::NEON, float>);
 #endif  // MACE_ENABLE_NEON
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("ResizeBilinear")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("ResizeBilinear")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     ResizeBilinearOp<DeviceType::OPENCL, float>);
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("ResizeBilinear")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("ResizeBilinear")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     ResizeBilinearOp<DeviceType::OPENCL, half>);
+}
 
 }  //  namespace mace

mace/ops/resize_bilinear_benchmark.cc

@@ -26,7 +26,8 @@ static void ResizeBilinearBenchmark(int iters,
   net.AddInputFromArray<D, index_t>("OutSize", {2},
                                     {output_height, output_width});
   if (D == DeviceType::OPENCL) {
-    BufferToImage<D, T>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+    BufferToImage<D, T>(net, "Input", "InputImage",
+                        kernels::BufferType::IN_OUT);
     OpDefBuilder("ResizeBilinear", "ResizeBilinearBenchmark")
         .Input("InputImage")
         .Input("OutSize")

mace/ops/resize_bilinear_test.cc

@@ -91,7 +91,8 @@ void TestRandomResizeBilinear() {
     expected.Copy(*net.GetOutput("Output"));
 
     if (D == DeviceType::OPENCL) {
-      BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+      BufferToImage<D, float>(net, "Input", "InputImage",
+                              kernels::BufferType::IN_OUT);
 
       OpDefBuilder("ResizeBilinear", "ResizeBilinearTest")
           .Input("InputImage")
@@ -102,7 +103,8 @@ void TestRandomResizeBilinear() {
       // Run
       net.RunOp(D);
 
-      ImageToBuffer<D, float>(net, "OutputImage", "DeviceOutput", kernels::BufferType::IN_OUT);
+      ImageToBuffer<D, float>(net, "OutputImage", "DeviceOutput",
+                              kernels::BufferType::IN_OUT);
     } else {
       // TODO support NEON
     }

mace/ops/space_to_batch.cc

@@ -6,13 +6,17 @@
 
 namespace mace {
 
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("SpaceToBatchND")
+void Register_SpaceToBatchND(OperatorRegistry *op_registry) {
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("SpaceToBatchND")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<float>("T")
                                      .Build(),
                     SpaceToBatchNDOp<DeviceType::OPENCL, float>);
-REGISTER_OPENCL_OPERATOR(OpKeyBuilder("SpaceToBatchND")
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("SpaceToBatchND")
+                                     .Device(DeviceType::OPENCL)
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     SpaceToBatchNDOp<DeviceType::OPENCL, half>);
+}
 
 }  //  namespace mace

mace/ops/space_to_batch_benchmark.cc

@@ -15,7 +15,8 @@ static void BMSpaceToBatch(
   OpsTestNet net;
   net.AddRandomInput<D, float>("Input", {batch, height, width, channels});
 
-  BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+  BufferToImage<D, float>(net, "Input", "InputImage",
+                          kernels::BufferType::IN_OUT);
   OpDefBuilder("SpaceToBatchND", "SpaceToBatchNDTest")
       .Input("InputImage")
       .Output("OutputImage")
@@ -37,14 +38,16 @@ static void BMSpaceToBatch(
 }
 
 #define BM_SPACE_TO_BATCH_MACRO(N, H, W, C, SHAPE, TYPE, DEVICE)             \
-  static void BM_SPACE_TO_BATCH_##N##_##H##_##W##_##C##_##SHAPE##_##TYPE##_##DEVICE( \
+  static void                                                                \
+      BM_SPACE_TO_BATCH_##N##_##H##_##W##_##C##_##SHAPE##_##TYPE##_##DEVICE( \
           int iters) {                                                       \
     const int64_t tot = static_cast<int64_t>(iters) * N * C * H * W;         \
     mace::testing::ItemsProcessed(tot);                                      \
     mace::testing::BytesProcessed(tot *(sizeof(TYPE)));                      \
     BMSpaceToBatch<DEVICE, TYPE>(iters, N, H, W, C, SHAPE);                  \
   }                                                                          \
-  BENCHMARK(BM_SPACE_TO_BATCH_##N##_##H##_##W##_##C##_##SHAPE##_##TYPE##_##DEVICE)
+  BENCHMARK(                                                                 \
+      BM_SPACE_TO_BATCH_##N##_##H##_##W##_##C##_##SHAPE##_##TYPE##_##DEVICE)
 
 #define BM_SPACE_TO_BATCH(N, H, W, C, SHAPE, TYPE) \
   BM_SPACE_TO_BATCH_MACRO(N, H, W, C, SHAPE, TYPE, OPENCL);

mace/ops/space_to_batch_test.cc

@@ -2,23 +2,23 @@
 // Copyright (c) 2017 XiaoMi All rights reserved.
 //
 
+#include <fstream>
 #include "gtest/gtest.h"
 #include "mace/ops/ops_test_util.h"
-#include <fstream>
 
 using namespace mace;
 
-template<DeviceType D>
+template <DeviceType D>
 void RunSpaceToBatch(const std::vector<index_t> &input_shape,
                      const std::vector<float> &input_data,
                      const std::vector<int> &block_shape_data,
                      const std::vector<int> &padding_data,
                      const Tensor *expected) {
   OpsTestNet net;
-  net.AddInputFromArray<D, float>(
-      "Input", input_shape, input_data);
+  net.AddInputFromArray<D, float>("Input", input_shape, input_data);
 
-  BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+  BufferToImage<D, float>(net, "Input", "InputImage",
+                          kernels::BufferType::IN_OUT);
   OpDefBuilder("SpaceToBatchND", "SpaceToBatchNDTest")
       .Input("InputImage")
       .Output("OutputImage")
@@ -29,12 +29,13 @@ void RunSpaceToBatch(const std::vector<index_t> &input_shape,
   // Run
   net.RunOp(D);
 
-  ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+  ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                          kernels::BufferType::IN_OUT);
   // Check
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-8);
 }
 
-template<DeviceType D>
+template <DeviceType D>
 void RunBatchToSpace(const std::vector<index_t> &input_shape,
                      const std::vector<float> &input_data,
                      const std::vector<int> &block_shape_data,
@@ -42,10 +43,10 @@ void RunBatchToSpace(const std::vector<index_t> &input_shape,
                      const Tensor *expected) {
   OpsTestNet net;
   // Add input data
-  net.AddInputFromArray<D, float>(
-      "Input", input_shape, input_data);
+  net.AddInputFromArray<D, float>("Input", input_shape, input_data);
 
-  BufferToImage<D, float>(net, "Input", "InputImage", kernels::BufferType::IN_OUT);
+  BufferToImage<D, float>(net, "Input", "InputImage",
+                          kernels::BufferType::IN_OUT);
   OpDefBuilder("BatchToSpaceND", "BatchToSpaceNDTest")
       .Input("InputImage")
       .Output("OutputImage")
@@ -56,21 +57,21 @@ void RunBatchToSpace(const std::vector<index_t> &input_shape,
   // Run
   net.RunOp(D);
 
-  ImageToBuffer<D, float>(net, "OutputImage", "Output", kernels::BufferType::IN_OUT);
+  ImageToBuffer<D, float>(net, "OutputImage", "Output",
+                          kernels::BufferType::IN_OUT);
   // Check
   ExpectTensorNear<float>(*expected, *net.GetOutput("Output"), 1e-8);
 }
 
-template<typename T>
+template <typename T>
 void TestBidirectionalTransform(const std::vector<index_t> &space_shape,
                                 const std::vector<float> &space_data,
                                 const std::vector<int> &block_data,
                                 const std::vector<int> &padding_data,
                                 const std::vector<index_t> &batch_shape,
                                 const std::vector<float> &batch_data) {
-
-  auto space_tensor = unique_ptr<Tensor>(new Tensor(GetDeviceAllocator(DeviceType::OPENCL),
-                                                    DataTypeToEnum<T>::v()));
+  auto space_tensor = unique_ptr<Tensor>(new Tensor(
+      GetDeviceAllocator(DeviceType::OPENCL), DataTypeToEnum<T>::v()));
   space_tensor->Resize(space_shape);
   {
     Tensor::MappingGuard space_mapper(space_tensor.get());
@@ -81,8 +82,8 @@ void TestBidirectionalTransform(const std::vector<index_t> &space_shape,
     memcpy(space_ptr, space_data.data(), space_data.size() * sizeof(T));
   }
 
-  auto batch_tensor = unique_ptr<Tensor>(new Tensor(GetDeviceAllocator(DeviceType::OPENCL),
-                                                    DataTypeToEnum<T>::v()));
+  auto batch_tensor = unique_ptr<Tensor>(new Tensor(
+      GetDeviceAllocator(DeviceType::OPENCL), DataTypeToEnum<T>::v()));
   batch_tensor->Resize(batch_shape);
   {
     Tensor::MappingGuard batch_mapper(batch_tensor.get());
@@ -91,113 +92,81 @@ void TestBidirectionalTransform(const std::vector<index_t> &space_shape,
     memcpy(batch_ptr, batch_data.data(), batch_data.size() * sizeof(T));
   }
 
-  RunSpaceToBatch<DeviceType::OPENCL>(space_shape, space_data,
-                                      block_data,
-                                      padding_data,
-                                      batch_tensor.get());
+  RunSpaceToBatch<DeviceType::OPENCL>(space_shape, space_data, block_data,
+                                      padding_data, batch_tensor.get());
 
-  RunBatchToSpace<DeviceType::OPENCL>(batch_shape, batch_data,
-                                      block_data,
-                                      padding_data,
-                                      space_tensor.get());
+  RunBatchToSpace<DeviceType::OPENCL>(batch_shape, batch_data, block_data,
+                                      padding_data, space_tensor.get());
 }
 
 TEST(SpaceToBatchTest, SmallData) {
-  TestBidirectionalTransform<float>({1, 2, 2, 1},
-                                    {1, 2, 3, 4},
-                                    {2, 2},
-                                    {0, 0, 0, 0},
-                                    {4, 1, 1, 1},
-                                    {1, 2, 3, 4}
-  );
+  TestBidirectionalTransform<float>({1, 2, 2, 1}, {1, 2, 3, 4}, {2, 2},
+                                    {0, 0, 0, 0}, {4, 1, 1, 1}, {1, 2, 3, 4});
 }
 
 TEST(SpaceToBatchTest, SmallDataWithOnePadding) {
-  TestBidirectionalTransform<float>({1, 2, 2, 1},
-                                    {1, 2, 3, 4},
-                                    {3, 3},
-                                    {1, 0, 1, 0},
-                                    {9, 1, 1, 1},
-                                    {0, 0, 0, 0, 1, 2, 0, 3, 4}
-  );
+  TestBidirectionalTransform<float>({1, 2, 2, 1}, {1, 2, 3, 4}, {3, 3},
+                                    {1, 0, 1, 0}, {9, 1, 1, 1},
+                                    {0, 0, 0, 0, 1, 2, 0, 3, 4});
 }
 
 TEST(SpaceToBatchTest, SmallDataWithTwoPadding) {
-  TestBidirectionalTransform<float>({1, 2, 2, 1},
-                                    {1, 2, 3, 4},
-                                    {2, 2},
-                                    {1, 1, 1, 1},
-                                    {4, 2, 2, 1},
-                                    {0, 0, 0, 4, 0, 0, 3, 0, 0, 2, 0, 0, 1, 0, 0, 0}
-  );
+  TestBidirectionalTransform<float>(
+      {1, 2, 2, 1}, {1, 2, 3, 4}, {2, 2}, {1, 1, 1, 1}, {4, 2, 2, 1},
+      {0, 0, 0, 4, 0, 0, 3, 0, 0, 2, 0, 0, 1, 0, 0, 0});
 }
 
 TEST(SpaceToBatchTest, SmallDataWithLargeImage) {
-  TestBidirectionalTransform<float>({1, 2, 10, 1},
-                                    {1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
-                                     11, 12, 13, 14, 15, 16, 17, 18, 19, 20},
-                                    {2, 2},
-                                    {0, 0, 0, 0},
-                                    {4, 1, 5, 1},
-                                    {1, 3, 5, 7, 9,
-                                     2, 4, 6, 8, 10,
-                                     11, 13, 15, 17, 19,
-                                     12, 14, 16, 18, 20}
-  );
+  TestBidirectionalTransform<float>(
+      {1, 2, 10, 1},
+      {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20},
+      {2, 2}, {0, 0, 0, 0}, {4, 1, 5, 1},
+      {1, 3, 5, 7, 9, 2, 4, 6, 8, 10, 11, 13, 15, 17, 19, 12, 14, 16, 18, 20});
 }
 
 TEST(SpaceToBatchTest, MultiChannelData) {
-  TestBidirectionalTransform<float>({1, 2, 2, 3},
-                                    {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12},
-                                    {2, 2},
-                                    {0, 0, 0, 0},
-                                    {4, 1, 1, 3},
-                                    {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}
-  );
+  TestBidirectionalTransform<float>(
+      {1, 2, 2, 3}, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}, {2, 2},
+      {0, 0, 0, 0}, {4, 1, 1, 3}, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12});
 }
 
 TEST(SpaceToBatchTest, LargerMultiChannelData) {
-  TestBidirectionalTransform<float>({1, 4, 4, 1},
-                                    {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
-                                    {2, 2},
-                                    {0, 0, 0, 0},
-                                    {4, 2, 2, 1},
-                                    {1, 3, 9, 11, 2, 4, 10, 12, 5, 7, 13, 15, 6, 8, 14, 16}
-  );
+  TestBidirectionalTransform<float>(
+      {1, 4, 4, 1}, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
+      {2, 2}, {0, 0, 0, 0}, {4, 2, 2, 1},
+      {1, 3, 9, 11, 2, 4, 10, 12, 5, 7, 13, 15, 6, 8, 14, 16});
 }
 
 TEST(SpaceToBatchTest, MultiBatchData) {
-  TestBidirectionalTransform<float>({2, 2, 4, 1},
-                                    {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
-                                    {2, 2},
-                                    {0, 0, 0, 0},
-                                    {8, 1, 2, 1},
-                                    {1, 3, 2, 4, 5, 7, 6, 8, 9, 11, 10, 12, 13, 15, 14, 16}
-  );
+  TestBidirectionalTransform<float>(
+      {2, 2, 4, 1}, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
+      {2, 2}, {0, 0, 0, 0}, {8, 1, 2, 1},
+      {1, 3, 2, 4, 5, 7, 6, 8, 9, 11, 10, 12, 13, 15, 14, 16});
 }
 
 TEST(SpaceToBatchTest, MultiBatchAndChannelData) {
-  TestBidirectionalTransform<float>({2, 2, 4, 2},
+  TestBidirectionalTransform<float>(
+      {2, 2, 4, 2},
       {1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15, 16,
        17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32},
-                                    {2, 2},
-                                    {0, 0, 0, 0},
-                                    {8, 1, 2, 2},
+      {2, 2}, {0, 0, 0, 0}, {8, 1, 2, 2},
       {1,  2,  5,  6,  3,  4,  7,  8,  9,  10, 13, 14, 11, 12, 15, 16,
-                                    17, 18, 21, 22, 19, 20, 23, 24, 25, 26, 29, 30, 27, 28, 31, 32}
-  );
+       17, 18, 21, 22, 19, 20, 23, 24, 25, 26, 29, 30, 27, 28, 31, 32});
 }
 
-//TEST(SpaceTobatchTest, CompareTF) {
+// TEST(SpaceTobatchTest, CompareTF) {
 //
 //  const std::string space_file = "/data/local/tmp/test/input";
 //  const std::string batch_file = "/data/local/tmp/test/output";
 //  const std::vector<index_t> space_shape = {1, 256, 256, 32};
-//  const int space_size = std::accumulate(space_shape.begin(), space_shape.end(), 1, std::multiplies<int>());
+//  const int space_size = std::accumulate(space_shape.begin(),
+//  space_shape.end(), 1, std::multiplies<int>());
 //  const std::vector<index_t> batch_shape = {4, 130, 130, 32};
-//  const int batch_size = std::accumulate(batch_shape.begin(), batch_shape.end(), 1, std::multiplies<int>());
+//  const int batch_size = std::accumulate(batch_shape.begin(),
+//  batch_shape.end(), 1, std::multiplies<int>());
 //
-//  auto space_tensor = unique_ptr<Tensor>(new Tensor(GetDeviceAllocator(DeviceType::OPENCL),
+//  auto space_tensor = unique_ptr<Tensor>(new
+//  Tensor(GetDeviceAllocator(DeviceType::OPENCL),
 //                                                    DataTypeToEnum<float>::v()));
 //  space_tensor->Resize(space_shape);
 //  std::vector<float> space_data(space_size, 0.0);
@@ -216,7 +185,8 @@ TEST(SpaceToBatchTest, MultiBatchAndChannelData) {
 //    VLOG(0) << "open space file failed";
 //  }
 //
-//  auto batch_tensor = unique_ptr<Tensor>(new Tensor(GetDeviceAllocator(DeviceType::OPENCL),
+//  auto batch_tensor = unique_ptr<Tensor>(new
+//  Tensor(GetDeviceAllocator(DeviceType::OPENCL),
 //                                                    DataTypeToEnum<float>::v()));
 //  std::vector<float> batch_data(batch_size, 0.0);
 //  batch_tensor->Resize(batch_shape);
@@ -231,7 +201,8 @@ TEST(SpaceToBatchTest, MultiBatchAndChannelData) {
 //    }
 //    Tensor::MappingGuard batch_mapper(batch_tensor.get());
 //    float *batch_ptr = batch_tensor->mutable_data<float>();
-//    MACE_CHECK(static_cast<size_t>(batch_tensor->size()) == batch_data.size());
+//    MACE_CHECK(static_cast<size_t>(batch_tensor->size()) ==
+//    batch_data.size());
 //    memcpy(batch_ptr, batch_data.data(), batch_data.size() * sizeof(float));
 //  }
 //
@@ -245,4 +216,3 @@ TEST(SpaceToBatchTest, MultiBatchAndChannelData) {
 //                                      {2, 2, 2, 2},
 //                                      space_tensor.get());
 //}
-

mace/python/tools/model.template

@@ -27,7 +27,7 @@ void Create{{tensor.name}}(std::vector<mace::ConstTensor> &tensors) {
 #include "mace/core/public/mace.h"
 
 namespace {
-static void UpdateOp(mace::OperatorDef &op,
+void UpdateOp(mace::OperatorDef &op,
               const std::string &name,
               const std::string &type,
               const std::vector<std::string> &inputs,

tools/bazel-adb-run.sh

@@ -17,9 +17,8 @@ BAZEL_BIN_PATH=${BAZEL_BIN_PATH#//}
 BAZEL_BIN_PATH=bazel-bin/$BAZEL_BIN_PATH
 BIN_NAME=`echo $BAZEL_TARGET | cut -d: -f2`
 
-ANDROID_ABI=armeabi-v7a
 ANDROID_ABI=arm64-v8a
-STRIP=""
+ANDROID_ABI=armeabi-v7a
 STRIP="--strip always"
 VLOG_LEVEL=0
 PROFILINE="--define profiling=true"
@@ -31,7 +30,7 @@ bazel build -c opt $STRIP --verbose_failures $BAZEL_TARGET \
    --crosstool_top=//external:android/crosstool \
    --host_crosstool_top=@bazel_tools//tools/cpp:toolchain  \
    --cpu=$ANDROID_ABI \
-   --define neon=true
+   --define neon=false
 
 if [ $? -ne 0 ]; then
   exit 1