Optimize the performance of C++ API

paddle/phi/api/include/tensor.h

@@ -324,7 +324,7 @@ class PADDLE_API Tensor final {
    *
    * @return std::shared_ptr<phi::TensorBase>
    */
-  std::shared_ptr<phi::TensorBase> impl() const;
+  const std::shared_ptr<phi::TensorBase>& impl() const;
 
   /**
    * @brief Set the implemention of current Tensor.
@@ -333,6 +333,13 @@ class PADDLE_API Tensor final {
    */
   void set_impl(const std::shared_ptr<phi::TensorBase>& impl);
 
+  /**
+   * @brief Set the implemention of current Tensor.
+   *
+   * @param impl
+   */
+  void set_impl(std::shared_ptr<phi::TensorBase>&& impl);
+
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
   /**
    * @brief Get the stream where the tensor is currently located

paddle/phi/api/lib/api_gen_utils.cc

@@ -95,12 +95,8 @@ paddle::optional<phi::MetaTensor> MakeMetaTensor(
 /* ------------------ for output ----------------------- */
 
 phi::DenseTensor* SetKernelOutput(Backend backend, Tensor* out) {
-  if (!out->initialized()) {
-    auto dense_tensor = std::make_shared<phi::DenseTensor>(
-        phi::make_intrusive<SharedStorage>(phi::TransToPhiPlace(backend)),
-        phi::DenseTensorMeta());
-    out->set_impl(dense_tensor);
-    return dense_tensor.get();
+  if (out->impl() == nullptr) {
+    out->set_impl(std::make_shared<phi::DenseTensor>());
   }
   return static_cast<phi::DenseTensor*>(out->impl().get());
 }
@@ -111,9 +107,7 @@ std::vector<phi::DenseTensor*> SetKernelOutput(size_t out_size,
   out->reserve(out_size);
   std::vector<phi::DenseTensor*> results(out_size);
   for (size_t i = 0; i < out_size; ++i) {
-    auto tensor_ptr = std::make_shared<phi::DenseTensor>(
-        phi::make_intrusive<SharedStorage>(phi::TransToPhiPlace(backend)),
-        phi::DenseTensorMeta());
+    auto tensor_ptr = std::make_shared<phi::DenseTensor>();
     results[i] = tensor_ptr.get();
     out->emplace_back();
     out->back().set_impl(tensor_ptr);

paddle/phi/api/lib/data_transform.cc

@@ -167,10 +167,7 @@ phi::DenseTensor TransformData(const phi::DenseTensor& tensor,
 
   if (NeedTransformPlace(
           out.place(), target_args_def.backend, transform_flag)) {
-    phi::DenseTensor result(
-        phi::make_intrusive<paddle::experimental::SharedStorage>(
-            phi::TransToPhiPlace(target_args_def.backend)),
-        {out.dtype(), out.dims(), out.layout()});
+    phi::DenseTensor result;
     framework::TransDataDevice(
         out, phi::TransToPhiPlace(target_args_def.backend), &result);
     out = result;
@@ -190,14 +187,14 @@ std::shared_ptr<phi::DenseTensor> PrepareData(
            tensor_in->dtype(), target_args_def.dtype, transform_flag) &&
        !NeedTransformLayout(
            tensor_in->layout(), target_args_def.layout, transform_flag))) {
-    return std::dynamic_pointer_cast<phi::DenseTensor>(tensor_in);
+    return std::static_pointer_cast<phi::DenseTensor>(tensor_in);
   }
 
   phi::DenseTensor out =
       TransformData(*(static_cast<phi::DenseTensor*>(tensor_in.get())),
                     target_args_def,
                     transform_flag);
-  return std::make_shared<phi::DenseTensor>(out);
+  return std::make_shared<phi::DenseTensor>(std::move(out));
 }
 
 std::shared_ptr<phi::DenseTensor> PrepareData(

paddle/phi/api/lib/tensor.cc

@@ -46,6 +46,7 @@ limitations under the License. */
  * In the future, the necessary components will be moved to the this library,
  * or the corresponding components will be re-implemented.
  */
+
 #include "paddle/fluid/memory/memory.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/platform/stream/cuda_stream.h"
@@ -142,7 +143,12 @@ PlaceType Tensor::place() const {
 }
 
 paddle::platform::Place Tensor::inner_place() const {
-  return ConvertExtPlaceToInnerPlace(place());
+  PADDLE_ENFORCE_NOT_NULL(
+      impl_,
+      phi::errors::PermissionDenied(
+          "Null pointer error, the impl_ of Tensor should not be "
+          "Null when calling Tensor::inner_place()."));
+  return impl_->place();
 }
 
 bool Tensor::is_cpu() const {
@@ -286,12 +292,16 @@ Tensor Tensor::slice(int64_t begin_idx, int64_t end_idx) const {
   }
 }
 
-std::shared_ptr<phi::TensorBase> Tensor::impl() const { return impl_; }
+const std::shared_ptr<phi::TensorBase> &Tensor::impl() const { return impl_; }
 
 void Tensor::set_impl(const std::shared_ptr<phi::TensorBase> &impl) {
   impl_ = impl;
 }
 
+void Tensor::set_impl(std::shared_ptr<phi::TensorBase> &&impl) {
+  impl_ = std::move(impl);
+}
+
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 gpuStream_t Tensor::stream() const {
   return platform::stream::get_current_stream(-1)->raw_stream();

paddle/phi/core/kernel_factory.h

@@ -197,8 +197,16 @@ class Kernel {
 
   const KernelArgsDef& args_def() const { return args_def_; }
 
+  const TensorArgDef& InputAt(size_t idx) const {
+    return args_def_.input_defs().at(idx);
+  }
+
   TensorArgDef& InputAt(size_t idx) { return args_def_.input_defs().at(idx); }
 
+  const TensorArgDef& OutputAt(size_t idx) const {
+    return args_def_.output_defs().at(idx);
+  }
+
   TensorArgDef& OutputAt(size_t idx) { return args_def_.output_defs().at(idx); }
 
   bool IsValid() { return fn_ != nullptr; }

python/paddle/utils/code_gen/api_base.py

@@ -698,7 +698,7 @@ def gen_dense_tensor_kernel_code(self, code_indent, inplace_flag=False):
             self.outputs['types'], 'SetKernelOutput', code_indent, inplace_flag)
         api_func_name = self.get_api_func_name() + ('_' if inplace_flag else '')
         return f"""
-{code_indent}  auto kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
+{code_indent}  const auto& kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
 {code_indent}      "{self.kernel['func'][0]}", {{kernel_backend, kernel_layout, kernel_data_type}});
 {code_indent}  VLOG(6) << "{self.api} API kernel key: [" << kernel_backend << ", " << kernel_layout << ", "<< kernel_data_type << "]";
 {code_indent}  VLOG(6) << "{self.api} API kernel: " << kernel;