Add Neon device

mace/core/allocator.h

@@ -12,8 +12,13 @@
 
 namespace mace {
 
-// 16 bytes = 32 * 4 (Neon)
+#ifdef __ANDROID__
+// 16 bytes = 128 bits = 32 * 4 (Neon)
 constexpr size_t kMaceAlignment = 16;
+#else
+// 32 bytes = 256 bits (AVX512)
+constexpr size_t kMaceAlignment = 32;
+#endif
 
 class Allocator {
  public:
@@ -41,25 +46,18 @@ class CPUAllocator: public Allocator {
     void* data = nullptr;
 #ifdef __ANDROID__
     data = memalign(kMaceAlignment, nbytes);
-#elif defined(_MSC_VER)
-    data = _aligned_malloc(nbytes, kMaceAlignment);
 #else
     CHECK(posix_memalign(&data, kMaceAlignment, nbytes) == 0);
 #endif
     CHECK_NOTNULL(data);
+    // TODO(heliangliang) This should be avoided sometimes
     memset(data, 0, nbytes);
     return data;
   }
 
-#ifdef _MSC_VER
-  void Delete(void* data) {
-    _aligned_free(data);
-  }
-#else
   void Delete(void* data) {
     free(data);
   }
-#endif
 
   void CopyBytes(void* dst, const void* src, size_t size) {
     memcpy(dst, src, size);
@@ -80,6 +78,11 @@ struct DeviceContext<DeviceType::CPU> {
   static Allocator* allocator() { return cpu_allocator(); }
 };
 
+template <>
+struct DeviceContext<DeviceType::NEON> {
+  static Allocator* allocator() { return cpu_allocator(); }
+};
+
 Allocator* GetDeviceAllocator(DeviceType type);
 
 } // namespace mace

mace/core/operator.cc

@@ -18,6 +18,13 @@ MACE_DEFINE_REGISTRY(
     Workspace*);
 MACE_REGISTER_DEVICE_TYPE(DeviceType::CPU, CPUOperatorRegistry);
 
+MACE_DEFINE_REGISTRY(
+    NEONOperatorRegistry,
+    OperatorBase,
+    const OperatorDef&,
+    Workspace*);
+MACE_REGISTER_DEVICE_TYPE(DeviceType::NEON, CPUOperatorRegistry);
+
 unique_ptr<OperatorBase> CreateOperator(
     const OperatorDef& operator_def,
     Workspace* ws,
@@ -33,4 +40,4 @@ OperatorBase::OperatorBase(const OperatorDef &operator_def, Workspace *ws)
 }
 
 
-} // namespace mace
\ No newline at end of file
+} // namespace mace

mace/core/operator.h

@@ -145,6 +145,17 @@ MACE_DECLARE_REGISTRY(
 #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__)
+
 unique_ptr<OperatorBase> CreateOperator(
     const OperatorDef &operator_def,
     Workspace *ws,

mace/ops/relu.cc

@@ -23,6 +23,23 @@ bool ReluOp<DeviceType::CPU, float>::Run() {
   return true;
 }
 
+template <>
+bool ReluOp<DeviceType::NEON, float>::Run() {
+  const Tensor* X = Input(0);
+  Tensor* Y = Output(0);
+  Y->ResizeLike(X);
+
+  const float* Xdata = X-> data<float>();
+  float* Ydata = Y->mutable_data<float>();
+  for (int i = 0; i < X->size(); ++i) {
+    Ydata[i] = std::max(Xdata[i], 0.f);
+    VLOG(0) << i << ": " << Xdata[i] << " " << Ydata[i];
+  }
+
+  return true;
+}
+
 REGISTER_CPU_OPERATOR(Relu, ReluOp<DeviceType::CPU, float>);
+REGISTER_NEON_OPERATOR(Relu, ReluOp<DeviceType::NEON, float>);
 
 } //  namespace mace

mace/proto/mace.proto

@@ -3,8 +3,9 @@ syntax = "proto2";
 package mace;
 
 enum DeviceType {
-  CPU = 0;                    // In default, we will use CPU.
-  GPU = 1;
+  CPU    = 0;                    // In default, we will use CPU.
+  NEON   = 1;
+  OPENCL = 2;
 }
 
 enum DataType {
@@ -70,4 +71,4 @@ message NetDef {
   optional string version = 3;
   repeated Argument arg = 4;
   repeated TensorProto tensors = 5;
-}
\ No newline at end of file
+}