add use_pinned

paddle/fluid/framework/tensor.h

@@ -45,10 +45,11 @@ class Tensor {
   friend struct EigenVector;
 
  public:
-  Tensor() : offset_(0) {}
+  Tensor() : offset_(0), use_pinned_(false) {}
 
   /*! Constructor with place should only be used in pybind. */
-  explicit Tensor(const platform::Place& place) : offset_(0) {
+  explicit Tensor(const platform::Place& place)
+      : offset_(0), use_pinned_(false) {
     holder_->set_place(place);
   }
 
@@ -69,11 +70,12 @@ class Tensor {
    * @note    If not exist, then allocation.
    */
   template <typename T>
-  inline T* mutable_data(platform::Place place);
+  inline T* mutable_data(platform::Place place, bool use_pinned = false);
 
-  inline void* mutable_data(platform::Place place, std::type_index type);
+  inline void* mutable_data(platform::Place place, std::type_index type,
+                            bool use_pinned = false);
 
-  inline void* mutable_data(platform::Place place);
+  inline void* mutable_data(platform::Place place, bool use_pinned = false);
 
   /**
    * @brief     Return a pointer to mutable memory block.
@@ -84,7 +86,8 @@ class Tensor {
    * @note      If not exist, then allocation.
    */
   template <typename T>
-  inline T* mutable_data(DDim dims, platform::Place place);
+  inline T* mutable_data(DDim dims, platform::Place place,
+                         bool use_pinned = false);
 
   /*! Return the dimensions of the memory block. */
   inline const DDim& dims() const;
@@ -92,6 +95,9 @@ class Tensor {
   /*! Return the numel of the memory block. */
   inline int64_t numel() const;
 
+  /*! Return the numel of the memory block. */
+  inline bool isPinned() const;
+
   /*! Resize the dimensions of the memory block. */
   inline Tensor& Resize(const DDim& dims);
 
@@ -146,12 +152,14 @@ class Tensor {
 
   template <typename Place>
   struct PlaceholderImpl : public Placeholder {
-    PlaceholderImpl(Place place, size_t size, std::type_index type)
-        : ptr_(static_cast<uint8_t*>(memory::Alloc(place, size)),
-               memory::PODDeleter<uint8_t, Place>(place)),
+    PlaceholderImpl(Place place, size_t size, std::type_index type,
+                    bool use_pinned = false)
+        : ptr_(static_cast<uint8_t*>(memory::Alloc(place, size, use_pinned)),
+               memory::PODDeleter<uint8_t, Place>(place, use_pinned)),
           place_(place),
           size_(size),
-          type_(type) {
+          type_(type),
+          use_pinned_(use_pinned) {
       PADDLE_ENFORCE_NOT_NULL(ptr_, "Insufficient %s memory to allocation.",
                               (is_cpu_place(place_) ? "CPU" : "GPU"));
     }
@@ -174,6 +182,9 @@ class Tensor {
 
     /* the current type of memory */
     std::type_index type_;
+
+    /*! use pinned memory or not. */
+    bool use_pinned_;
   };
 
   /*! holds the memory block if allocated. */
@@ -208,6 +219,7 @@ class Tensor {
    *          PlaceHolder::ptr_ and where the tensor data really begins.
    */
   size_t offset_;
+  bool use_pinned_;
 };
 
 inline void Tensor::switch_place(platform::Place new_place) {

paddle/fluid/framework/tensor_impl.h

@@ -101,19 +101,21 @@ inline T* Tensor::data() {
 }
 
 template <typename T>
-inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
+inline T* Tensor::mutable_data(DDim dims, platform::Place place,
+                               bool use_pinned) {
   static_assert(std::is_pod<T>::value, "T must be POD");
   Resize(dims);
-  return mutable_data<T>(place);
+  return mutable_data<T>(place, use_pinned);
 }
 
 template <typename T>
-inline T* Tensor::mutable_data(platform::Place place) {
+inline T* Tensor::mutable_data(platform::Place place, bool use_pinned) {
   static_assert(std::is_pod<T>::value, "T must be POD");
-  return reinterpret_cast<T*>(mutable_data(place, typeid(T)));
+  return reinterpret_cast<T*>(mutable_data(place, typeid(T), use_pinned));
 }
 
-inline void* Tensor::mutable_data(platform::Place place, std::type_index type) {
+inline void* Tensor::mutable_data(platform::Place place, std::type_index type,
+                                  bool use_pinned) {
   if (holder_ != nullptr) {
     holder_->set_type(type);
   }
@@ -127,26 +129,27 @@ inline void* Tensor::mutable_data(platform::Place place, std::type_index type) {
       holder_->size() < size + offset_) {
     if (platform::is_cpu_place(place)) {
       holder_.reset(new PlaceholderImpl<platform::CPUPlace>(
-          boost::get<platform::CPUPlace>(place), size, type));
+          boost::get<platform::CPUPlace>(place), size, type, use_pinned));
     } else if (platform::is_gpu_place(place)) {
 #ifndef PADDLE_WITH_CUDA
       PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
     }
 #else
       holder_.reset(new PlaceholderImpl<platform::CUDAPlace>(
-          boost::get<platform::CUDAPlace>(place), size, type));
+          boost::get<platform::CUDAPlace>(place), size, type, use_pinned));
     }
 #endif
     offset_ = 0;
+    use_pinned_ = use_pinned;
   }
   return reinterpret_cast<void*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
                                  offset_);
 }
 
-inline void* Tensor::mutable_data(platform::Place place) {
+inline void* Tensor::mutable_data(platform::Place place, bool use_pinned) {
   PADDLE_ENFORCE(this->holder_ != nullptr,
                  "Cannot invoke mutable data if current hold nothing");
-  return mutable_data(place, holder_->type());
+  return mutable_data(place, holder_->type(), use_pinned);
 }
 
 inline Tensor& Tensor::ShareDataWith(const Tensor& src) {
@@ -188,6 +191,8 @@ inline const DDim& Tensor::dims() const { return dims_; }
 
 inline int64_t Tensor::numel() const { return product(dims_); }
 
+inline bool Tensor::isPinned() const { return use_pinned_; }
+
 inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
   Tensor res;
   res.ShareDataWith(src);