Merge branch 'master' into 'master'

AddN op; Refactor op default template impl.

See merge request !11

mace/examples/helloworld.cc

@@ -27,10 +27,11 @@ int main() {
   arg_1->set_f(1.5);
 
   OperatorDef op_def_2;
+  op_def_2.add_input("Output0");
   op_def_2.add_input("Output1");
   op_def_2.add_output("Output2");
-  op_def_2.set_name("ReluTest2");
-  op_def_2.set_type("Relu");
+  op_def_2.set_name("AddNTest");
+  op_def_2.set_type("AddN");
   auto arg_2 = op_def_2.add_arg();
   arg_2->set_name("arg0");
   arg_2->set_f(2.5);

mace/kernels/addn.h

@@ -18,10 +18,10 @@ void AddNFuntion(const vector<const Tensor*>& input_tensor, Tensor *output_tenso
   int64 size = input_tensor[0]->size();
   vector<const T*> inputs(n);
   for (int i = 0; i < n; ++i) {
-    inputs[i] = input_tensor[i]->data<float>();
+    inputs[i] = input_tensor[i]->data<T>();
   }
   output_tensor->ResizeLike(input_tensor[0]);
-  float* output = output_tensor->mutable_data<T>();
+  T* output = output_tensor->mutable_data<T>();
 
   for (int i = 0; i < n; ++i) {
     for (int64 j = 0; j < size; ++j) {

mace/kernels/relu.h

@@ -14,8 +14,8 @@ template<typename T>
 void ReluFuntion(const Tensor *input_tensor, Tensor *output_tensor) {
   int64 size = input_tensor->size();
   output_tensor->ResizeLike(input_tensor);
-  const float *input = input_tensor->data<float>();
-  float *output = output_tensor->mutable_data<float>();
+  const T *input = input_tensor->data<T>();
+  T *output = output_tensor->mutable_data<T>();
 
   for (int64 i = 0; i < size; ++i) {
     output[i] = std::max(input[i], static_cast<T>(0));

mace/ops/BUILD

@@ -12,7 +12,7 @@ load("//mace:mace.bzl", "if_android")
 
 cc_library(
     name = "ops",
-    srcs = ["relu.cc"],
+    srcs = glob(["*.cc"]),
     hdrs = glob(["*.h"]),
     deps = [
         "//mace/proto:cc_proto",
@@ -23,19 +23,4 @@ cc_library(
     alwayslink = 1,
 )
 
-cc_test(
-    name = "relu_test",
-    srcs = ["relu_test.cc",],
-    deps = [
-        "@gtest//:gtest_main",
-        ":ops",
-    ],
-    copts = ['-std=c++11'],
-    linkopts = if_android([
-        "-pie",
-        "-llog",
-        "-latomic",
-        ]),
-    linkstatic = 1,
-)
 

mace/ops/addn.cc

+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#include "mace/ops/addn.h"
+#include "mace/proto/mace.pb.h"
+#if __ARM_NEON
+#include "mace/kernels/neon/addn_neon.h"
+#endif // __ARM_NEON
+
+namespace mace {
+
+REGISTER_CPU_OPERATOR(AddN, AddNOp<DeviceType::CPU, float>);
+
+#if __ARM_NEON
+template <>
+bool AddNOp<DeviceType::NEON, float>::Run() {
+  Tensor* output_tensor = Output(0);
+  kernels::NeonAddNFuntion_float(Inputs(), output_tensor);
+  return true;
+}
+REGISTER_NEON_OPERATOR(AddN, AddNOp<DeviceType::NEON, float>);
+#endif // __ARM_NEON
+
+} //  namespace mace

mace/ops/addn.h

+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#ifndef MACE_OPS_ADDN_H_
+#define MACE_OPS_ADDN_H_
+
+#include "mace/core/operator.h"
+#include "mace/kernels/addn.h"
+
+namespace mace {
+
+template<DeviceType D, class T>
+class AddNOp : public Operator<D, T> {
+ public:
+  AddNOp(const OperatorDef &operator_def, Workspace *ws)
+      : Operator<D, T>(operator_def, ws) {}
+  bool Run() override {
+    Tensor* output_tensor = this->Output(0);
+    kernels::AddNFuntion<T>(this->Inputs(), output_tensor);
+    return true;
+  }
+};
+
+} // namespace mace
+
+#endif // MACE_OPS_ADDN_H_

mace/ops/relu.cc

@@ -4,23 +4,14 @@
 
 #include "mace/ops/relu.h"
 #include "mace/proto/mace.pb.h"
-#include "mace/kernels/relu.h"
 #if __ARM_NEON
 #include "mace/kernels/neon/relu_neon.h"
 #endif // __ARM_NEON
 
 namespace mace {
 
-template <>
-bool ReluOp<DeviceType::CPU, float>::Run() {
-  const Tensor* input_tensor = Input(0);
-  Tensor* output_tensor = Output(0);
-  kernels::ReluFuntion<float>(input_tensor, output_tensor);
-  return true;
-}
 REGISTER_CPU_OPERATOR(Relu, ReluOp<DeviceType::CPU, float>);
 
-
 #if __ARM_NEON
 template <>
 bool ReluOp<DeviceType::NEON, float>::Run() {

mace/ops/relu.h

@@ -6,6 +6,7 @@
 #define MACE_OPS_RELU_H_
 
 #include "mace/core/operator.h"
+#include "mace/kernels/relu.h"
 
 namespace mace {
 
@@ -14,7 +15,12 @@ class ReluOp : public Operator<D, T> {
  public:
   ReluOp(const OperatorDef &operator_def, Workspace *ws)
       : Operator<D, T>(operator_def, ws) {}
-  bool Run() override;
+  bool Run() override {
+    const Tensor* input_tensor = this->Input(0);
+    Tensor* output_tensor = this->Output(0);
+    kernels::ReluFuntion<T>(input_tensor, output_tensor);
+    return true;
+  }
 };
 
 } // namespace mace

mace/ops/relu_test.cc

-//
-// Copyright (c) 2017 XiaoMi All rights reserved.
-//
-
-#include "gtest/gtest.h"
-
-#include "mace/core/operator.h"
-#include "mace/core/net.h"
-
-using namespace mace;
-
-TEST(ReluTest, Relu) {
-  OperatorRegistry* registry = gDeviceTypeRegistry()->at(DeviceType::CPU);
-  vector<string> registry_keys = registry->Keys();
-  for (auto& key: registry_keys) {
-    VLOG(0) << "registry_op: " << key;
-  }
-
-  // Construct graph
-  OperatorDef op_def;
-  op_def.add_input("Input0");
-  op_def.add_output("Output0");
-  op_def.set_name("ReluTest");
-  op_def.set_type("Relu");
-  auto arg = op_def.add_arg();
-  arg->set_name("arg0");
-  arg->set_f(1.5);
-
-  NetDef net_def;
-  net_def.set_name("NetTest");
-  net_def.add_op()->CopyFrom(op_def);
-
-  VLOG(0) << net_def.DebugString();
-
-  // Create workspace and input tensor
-  Workspace ws;
-  Tensor* input = ws.CreateTensor("Input0", cpu_allocator(), DataType::DT_FLOAT);
-  input->Resize({2,3});
-  float* input_data = input->mutable_data<float>();
-  for (int i = 0; i < 6; ++i) {
-    input_data[i] = i-3;
-  }
-
-  // Create Net & run
-  auto net = CreateNet(net_def, &ws, DeviceType::CPU);
-  net->Run();
-
-  // Create Op & run
-  auto op = CreateOperator(op_def, &ws, DeviceType::CPU);
-  ASSERT_FLOAT_EQ(1.5f, op->GetSingleArgument<float>("arg0", 1.0f));
-
-}
\ No newline at end of file