Fix MixedVector

paddle/fluid/framework/details/cow_ptr.h

@@ -28,31 +28,27 @@ class COWPtr {
  private:
   RefPtr m_sp;
 
-  void detach() {
-    T* tmp = m_sp.get();
-    if (!(tmp == nullptr || m_sp.unique())) {
-      m_sp = RefPtr(new T(*tmp));
-    }
-  }
-
  public:
   COWPtr() : m_sp(nullptr) {}
   explicit COWPtr(T* t) : m_sp(t) {}
-  explicit COWPtr(const RefPtr& refptr) : m_sp(refptr) {}
 
-  const T& Data() const { return operator*(); }
+  const T& Data() const { return *m_sp; }
 
-  T* MutableData() { return operator->(); }
+  T* MutableData() {
+    DetachIfNotUnique();
+    return m_sp.get();
+  }
 
-  const T& operator*() const { return *m_sp; }
-  T& operator*() {
-    detach();
-    return *m_sp;
+  void DetachIfNotUnique() {
+    T* tmp = m_sp.get();
+    if (!(tmp == nullptr || m_sp.unique())) {
+      Detach();
+    }
   }
-  const T* operator->() const { return m_sp.operator->(); }
-  T* operator->() {
-    detach();
-    return m_sp.operator->();
+
+  void Detach() {
+    T* tmp = m_sp.get();
+    m_sp = RefPtr(new T(*tmp));
   }
 };
 }  // namespace details

paddle/fluid/operators/lookup_table_op.cu

@@ -127,10 +127,8 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
       auto gpu_place = boost::get<platform::CUDAPlace>(context.GetPlace());
 
       // TODO(yuyang18): Strange code here.
-      memory::Copy(platform::CPUPlace(),
-                   new_rows.CUDAMutableData(context.GetPlace()), gpu_place,
-                   ids_data, ids_num * sizeof(int64_t), stream);
-
+      memory::Copy(gpu_place, new_rows.CUDAMutableData(context.GetPlace()),
+                   gpu_place, ids_data, ids_num * sizeof(int64_t), stream);
       d_table->set_rows(new_rows);
 
       auto *d_table_value = d_table->mutable_value();

paddle/fluid/operators/sgd_op.cu

@@ -89,7 +89,7 @@ class SGDOpCUDAKernel : public framework::OpKernel<T> {
       PADDLE_ENFORCE_EQ(in_height, out_dims[0]);
 
       auto& in_value = grad->value();
-      framework::Vector<int64_t> in_rows(grad->rows());
+      auto& in_rows = grad->rows();
 
       int64_t in_row_numel = in_value.numel() / in_rows.size();
       PADDLE_ENFORCE_EQ(in_row_numel, param_out->numel() / in_height);