Adding implementation for copying a vector to a tensor (#4635)

* Adding implementation for copying a vector to tensor
* Changing Tensor test to access gpu memory indirectly

paddle/framework/tensor.h

@@ -95,6 +95,19 @@ class Tensor {
   template <typename T>
   inline void CopyFrom(const Tensor& src, const platform::Place& dst_place);
 
+  /**
+   * @brief   Copy the content of an external vector to a tensor.
+   *
+   * @param[in] src   The external vector.
+   * @param[in] ctx   The device context contains place where to store.
+   *
+   * * @note    CopyFromVector assumes that the tensor has been resized
+   *            before invoking.
+   */
+  template <typename T>
+  inline void CopyFromVector(const std::vector<T>& src,
+                             const platform::Place& dst_place);
+
   /**
    * @brief   Return the slice of the tensor.
    *

paddle/framework/tensor_impl.h

@@ -123,6 +123,29 @@ inline void Tensor::CopyFrom(const Tensor& src,
 #endif
 }
 
+template <typename T>
+inline void Tensor::CopyFromVector(const std::vector<T>& src,
+                                   const platform::Place& dst_place) {
+  auto src_ptr = static_cast<const void*>(src.data());
+  platform::CPUPlace src_place;
+  auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
+  auto size = src.size() * sizeof(T);
+
+  if (platform::is_cpu_place(dst_place)) {
+    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr, src_place,
+                 src_ptr, size);
+  }
+#ifdef PADDLE_WITH_CUDA
+  else if (platform::is_gpu_place(dst_place)) {
+    memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr, src_place,
+                 src_ptr, size, 0);
+  }
+  PADDLE_ENFORCE(cudaStreamSynchronize(0),
+                 "cudaStreamSynchronize failed in Tensor CopyFromVector");
+
+#endif
+}
+
 template <typename T>
 inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
   check_memory_size<T>();

paddle/framework/tensor_test.cc

@@ -263,6 +263,93 @@ TEST(Tensor, CopyFrom) {
 #endif
 }
 
+TEST(Tensor, CopyFromVector) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  {
+    std::vector<int> src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9};
+    Tensor cpu_tensor;
+
+    // Copy to CPU Tensor
+    cpu_tensor.Resize(make_ddim({3, 3}));
+    auto cpu_place = new paddle::platform::CPUPlace();
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+
+    // Compare Tensors
+    const int* cpu_ptr = cpu_tensor.data<int>();
+    const int* src_ptr = src_vec.data();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+    }
+
+    src_vec.erase(src_vec.begin(), src_vec.begin() + 5);
+    cpu_tensor.Resize(make_ddim({2, 2}));
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+    cpu_ptr = cpu_tensor.data<int>();
+    src_ptr = src_vec.data();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    for (size_t i = 0; i < 5; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+    }
+
+    delete cpu_place;
+  }
+
+#ifdef PADDLE_WITH_CUDA
+  {
+    std::vector<int> src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9};
+    Tensor cpu_tensor;
+    Tensor gpu_tensor;
+    Tensor dst_tensor;
+
+    // Copy to CPU Tensor
+    cpu_tensor.Resize(make_ddim({3, 3}));
+    auto cpu_place = new paddle::platform::CPUPlace();
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+
+    // Copy to GPUTensor
+    gpu_tensor.Resize(make_ddim({3, 3}));
+    auto gpu_place = new paddle::platform::GPUPlace();
+    gpu_tensor.CopyFromVector<int>(src_vec, *gpu_place);
+    // Copy from GPU to CPU tensor for comparison
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+
+    // Compare Tensors
+    const int* src_ptr = src_vec.data();
+    const int* cpu_ptr = cpu_tensor.data<int>();
+    const int* dst_ptr = dst_tensor.data<int>();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    ASSERT_NE(src_ptr, dst_ptr);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+      EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+    }
+
+    src_vec.erase(src_vec.begin(), src_vec.begin() + 5);
+
+    cpu_tensor.Resize(make_ddim({2, 2}));
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+    gpu_tensor.Resize(make_ddim({2, 2}));
+    gpu_tensor.CopyFromVector<int>(src_vec, *gpu_place);
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+
+    src_ptr = src_vec.data();
+    cpu_ptr = cpu_tensor.data<int>();
+    dst_ptr = dst_tensor.data<int>();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    ASSERT_NE(src_ptr, dst_ptr);
+    for (size_t i = 0; i < 5; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+      EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+    }
+
+    delete cpu_place;
+    delete gpu_place;
+  }
+#endif
+}
+
 TEST(Tensor, ReshapeToMatrix) {
   using namespace paddle::framework;
   using namespace paddle::platform;