Merge pull request #2809 from Canpio/dev_add_tensor_interface

add tensor interfaces

paddle/framework/tensor.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <cstdint>
 #include <memory>
 #include <type_traits>
 #include "paddle/framework/ddim.h"
@@ -26,31 +27,65 @@ namespace framework {
 
 class Tensor {
  public:
+  Tensor() : offset_(0) {}
+
+  explicit Tensor(const DDim& dims) : dims_(dims), offset_(0) {}
+
   template <typename T>
   const T* data() const {
-    PADDLE_ENFORCE(holder_ != nullptr,
-                   "Tensor::data must be called after Tensor::mutable_data.");
-    return static_cast<const T*>(holder_->Ptr());
+    PADDLE_ENFORCE(
+        holder_ != nullptr,
+        "Tenosr has not been initialized. Call Tensor::mutable_data first.");
+    return reinterpret_cast<const T*>(
+        reinterpret_cast<uintptr_t>(holder_->Ptr()) + offset_);
   }
 
   template <typename T,  // must be POD types
             typename std::enable_if<std::is_pod<T>::value>::type* = nullptr>
   T* mutable_data(DDim dims, paddle::platform::Place place) {
+    dims_ = dims;
     if (holder_ == nullptr ||
         !(holder_->Place() ==
           place) /* some versions of boost::variant don't have operator!= */
-        || holder_->Size() < product(dims) * sizeof(T)) {
+        || holder_->Size() < product(dims) * sizeof(T) + offset_) {
       holder_.reset(new PlaceholderImpl<T>(place, product(dims) * sizeof(T)));
+      offset_ = 0;
     }
-    return static_cast<T*>(holder_->Ptr());
+    return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->Ptr()) +
+                                offset_);
   }
 
-  template <typename T,  // must be POD types
-            typename std::enable_if<std::is_pod<T>::value>::type* = nullptr>
-  T* mutable_data(DDim dims) {
-    return mutable_data<T>(dims, paddle::platform::get_place());
+  void ShareDataFrom(const Tensor& src) {
+    PADDLE_ENFORCE(src.holder_ != nullptr,
+                   "Can not share data from an uninitialized tensor.");
+    holder_ = src.holder_;
+    dims_ = src.dims_;
+    offset_ = src.offset_;
   }
 
+  Tensor Slice(const int& begin_idx, const int& end_idx) const {
+    PADDLE_ENFORCE(holder_ != nullptr,
+                   "The sliced tenosr has not been initialized.");
+    PADDLE_ENFORCE(begin_idx >= 0 && end_idx <= dims_[0],
+                   "Slice index is less than zero or out of bound.");
+    PADDLE_ENFORCE(begin_idx < end_idx,
+                   "Begin index must be less than end index.");
+    PADDLE_ENFORCE(dims_[0] != 1, "Can not slice a tensor with dims_[0] = 1.");
+    std::vector<int> d = vectorize(dims_);
+    int base = 1;
+    for (size_t i = 1; i < d.size(); ++i) {
+      base *= d[i];
+    }
+    Tensor dst;
+    dst.holder_ = holder_;
+    dst.dims_ = dims_;
+    dst.dims_[0] = end_idx - begin_idx;
+    dst.offset_ = offset_ + begin_idx * base * holder_->TypeSize();
+    return dst;
+  }
+
+  DDim dims() const { return dims_; }
+
  private:
   // Placeholder hides type T, so it doesn't appear as a template
   // parameter of Variable.
@@ -59,6 +94,7 @@ class Tensor {
     virtual void* Ptr() const = 0;
     virtual paddle::platform::Place Place() const = 0;
     virtual size_t Size() const = 0;
+    virtual size_t TypeSize() const = 0;
   };
 
   template <typename T>
@@ -85,6 +121,7 @@ class Tensor {
     virtual void* Ptr() const { return static_cast<void*>(ptr_.get()); }
     virtual size_t Size() const { return size_; }
     virtual paddle::platform::Place Place() const { return place_; }
+    virtual size_t TypeSize() const { return sizeof(T); }
 
     std::unique_ptr<T, Deleter> ptr_;
     paddle::platform::Place place_;  // record the place of ptr_.
@@ -92,6 +129,8 @@ class Tensor {
   };
 
   std::shared_ptr<Placeholder> holder_;  // holds the memory block if allocated.
+  DDim dims_;
+  size_t offset_;  // marks the begin of tensor data area.
 };
 
 }  // namespace framework

paddle/framework/tensor_test.cc

@@ -15,15 +15,27 @@
 #include <gtest/gtest.h>
 #include <string>
 
-TEST(Tensor, ASSERT) {
-  paddle::framework::Tensor cpu_tensor;
+TEST(Tensor, Dims) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  Tensor tt(make_ddim({2, 3, 4}));
+  DDim dims = tt.dims();
+  ASSERT_EQ(arity(dims), 3);
+  for (int i = 0; i < 3; ++i) {
+    EXPECT_EQ(i + 2, dims[i]);
+  }
+}
+
+TEST(Tensor, DataAssert) {
+  paddle::framework::Tensor src_tensor;
 
   bool caught = false;
   try {
-    const double* p __attribute__((unused)) = cpu_tensor.data<double>();
+    src_tensor.data<double>();
   } catch (paddle::framework::EnforceNotMet err) {
     caught = true;
-    std::string msg = "Tensor::data must be called after Tensor::mutable_data.";
+    std::string msg =
+        "Tenosr has not been initialized. Call Tensor::mutable_data first.";
     const char* what = err.what();
     for (size_t i = 0; i < msg.length(); ++i) {
       ASSERT_EQ(what[i], msg[i]);
@@ -32,54 +44,138 @@ TEST(Tensor, ASSERT) {
   ASSERT_TRUE(caught);
 }
 
-/* mutable_data() is not tested at present
+/* following tests are not available at present
    because Memory::Alloc() and Memory::Free() have not been ready.
 
 TEST(Tensor, MutableData) {
   using namespace paddle::framework;
   using namespace paddle::platform;
   {
-    Tensor cpu_tensor;
+    Tensor src_tensor;
     float* p1 = nullptr;
     float* p2 = nullptr;
     // initialization
-    p1 = cpu_tensor.mutable_data<float>(make_ddim({1, 2, 3}), CPUPlace());
+    p1 = src_tensor.mutable_data<float>(make_ddim({1, 2, 3}), CPUPlace());
     EXPECT_NE(p1, nullptr);
-    // set cpu_tensor a new dim with large size
+    // set src_tensor a new dim with large size
     // momery is supposed to be re-allocated
-    p2 = cpu_tensor.mutable_data<float>(make_ddim({3, 4}));
+    p2 = src_tensor.mutable_data<float>(make_ddim({3, 4}), CPUPlace());
     EXPECT_NE(p2, nullptr);
     EXPECT_NE(p1, p2);
-    // set cpu_tensor a new dim with same size
+    // set src_tensor a new dim with same size
     // momery block is supposed to be unchanged
-    p1 = cpu_tensor.mutable_data<float>(make_ddim({2, 2, 3}));
+    p1 = src_tensor.mutable_data<float>(make_ddim({2, 2, 3}), CPUPlace());
     EXPECT_EQ(p1, p2);
-    // set cpu_tensor a new dim with smaller size
+    // set src_tensor a new dim with smaller size
     // momery block is supposed to be unchanged
-    p2 = cpu_tensor.mutable_data<float>(make_ddim({2, 2}));
+    p2 = src_tensor.mutable_data<float>(make_ddim({2, 2}), CPUPlace());
     EXPECT_EQ(p1, p2);
   }
 
   {
-    Tensor gpu_tensor;
+    Tensor src_tensor;
     float* p1 = nullptr;
     float* p2 = nullptr;
     // initialization
-    p1 = gpu_tensor.mutable_data<float>(make_ddim({1, 2, 3}), GPUPlace());
+    p1 = src_tensor.mutable_data<float>(make_ddim({1, 2, 3}), GPUPlace());
     EXPECT_NE(p1, nullptr);
-    // set gpu_tensor a new dim with large size
+    // set src_tensor a new dim with large size
     // momery is supposed to be re-allocated
-    p2 = gpu_tensor.mutable_data<float>(make_ddim({3, 4}));
+    p2 = src_tensor.mutable_data<float>(make_ddim({3, 4}), GPUPlace());
     EXPECT_NE(p2, nullptr);
     EXPECT_NE(p1, p2);
-    // set gpu_tensor a new dim with same size
+    // set src_tensor a new dim with same size
     // momery block is supposed to be unchanged
-    p1 = gpu_tensor.mutable_data<float>(make_ddim({2, 2, 3}));
+    p1 = src_tensor.mutable_data<float>(make_ddim({2, 2, 3}), GPUPlace());
     EXPECT_EQ(p1, p2);
-    // set gpu_tensor a new dim with smaller size
+    // set src_tensor a new dim with smaller size
     // momery block is supposed to be unchanged
-    p2 = gpu_tensor.mutable_data<float>(make_ddim({2, 2}));
+    p2 = src_tensor.mutable_data<float>(make_ddim({2, 2}), GPUPlace());
     EXPECT_EQ(p1, p2);
   }
 }
+
+TEST(Tensor, ShareDataFrom) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  {
+    Tensor src_tensor;
+    Tensor dst_tensor;
+    // Try to share data form uninitialized tensor
+    bool caught = false;
+    try {
+      dst_tensor.ShareDataFrom(src_tensor);
+    } catch (EnforceNotMet err) {
+      caught = true;
+      std::string msg = "Can not share data from an uninitialized tensor.";
+      const char* what = err.what();
+      for (size_t i = 0; i < msg.length(); ++i) {
+        ASSERT_EQ(what[i], msg[i]);
+      }
+    }
+    ASSERT_TRUE(caught);
+
+    src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), CPUPlace());
+    dst_tensor.ShareDataFrom(src_tensor);
+    ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
+  }
+
+  {
+    Tensor src_tensor;
+    Tensor dst_tensor;
+    src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), GPUPlace());
+    dst_tensor.ShareDataFrom(src_tensor);
+    ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
+  }
+}
+
+TEST(Tensor, Slice) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  {
+    Tensor src_tensor;
+    src_tensor.mutable_data<int>(make_ddim({5, 3, 4}), CPUPlace());
+    Tensor slice_tensor = src_tensor.Slice(1, 3);
+    DDim slice_dims = slice_tensor.dims();
+    ASSERT_EQ(arity(slice_dims), 3);
+    EXPECT_EQ(slice_dims[0], 2);
+    EXPECT_EQ(slice_dims[1], 3);
+    EXPECT_EQ(slice_dims[2], 4);
+
+    uintptr_t src_data_address =
+        reinterpret_cast<uintptr_t>(src_tensor.data<int>());
+    uintptr_t src_mutable_data_address = reinterpret_cast<uintptr_t>(
+        src_tensor.mutable_data<int>(src_tensor.dims(), CPUPlace()));
+    uintptr_t slice_data_address =
+        reinterpret_cast<uintptr_t>(slice_tensor.data<int>());
+    uintptr_t slice_mutable_data_address = reinterpret_cast<uintptr_t>(
+        slice_tensor.mutable_data<int>(slice_tensor.dims(), CPUPlace()));
+    EXPECT_EQ(src_data_address, src_mutable_data_address);
+    EXPECT_EQ(slice_data_address, slice_mutable_data_address);
+    EXPECT_EQ(src_data_address + 3 * 4 * 1 * sizeof(int), slice_data_address);
+  }
+
+  {
+    Tensor src_tensor;
+    src_tensor.mutable_data<double>(make_ddim({6, 9}), GPUPlace());
+    Tensor slice_tensor = src_tensor.Slice(2, 6);
+    DDim slice_dims = slice_tensor.dims();
+    ASSERT_EQ(arity(slice_dims), 2);
+    EXPECT_EQ(slice_dims[0], 4);
+    EXPECT_EQ(slice_dims[1], 9);
+
+    uintptr_t src_data_address =
+        reinterpret_cast<uintptr_t>(src_tensor.data<double>());
+    uintptr_t src_mutable_data_address = reinterpret_cast<uintptr_t>(
+        src_tensor.mutable_data<double>(src_tensor.dims(), GPUPlace()));
+    uintptr_t slice_data_address =
+        reinterpret_cast<uintptr_t>(slice_tensor.data<double>());
+    uintptr_t slice_mutable_data_address = reinterpret_cast<uintptr_t>(
+        slice_tensor.mutable_data<double>(slice_tensor.dims(), GPUPlace()));
+    EXPECT_EQ(src_data_address, src_mutable_data_address);
+    EXPECT_EQ(slice_data_address, slice_mutable_data_address);
+    EXPECT_EQ(src_data_address + 9 * 2 * sizeof(double), slice_data_address);
+  }
+}
+
 */
\ No newline at end of file