[PTen] Remove offset in storage (#38472)

* remove offset in storage

* revert api change

* fix custom op slice bug

* fix mutable_data error

paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.cc

@@ -83,7 +83,7 @@ void ScaleAPI(const egr::EagerTensor& x, float scale, float bias,
                       SizeOf(dense_tensor->dtype());
   auto dense_out = std::make_shared<pten::DenseTensor>(
       pten::make_intrusive<paddle::experimental::SharedStorage>(
-          paddle::memory::Alloc(place, bytes_size), 0),
+          paddle::memory::Alloc(place, bytes_size)),
       std::move(tensor_meta));
   // Handle Device Context
   const paddle::platform::Place& expected_kernel_place =

paddle/fluid/eager/tests/task_tests/fwd_bwd_joint_test.cc

@@ -41,7 +41,7 @@ egr::EagerTensor hook_function(const egr::EagerTensor& t) {
       paddle::framework::product(t_dense->dims()) * SizeOf(t_dense->dtype());
   auto ret_dense = std::make_shared<pten::DenseTensor>(
       pten::make_intrusive<paddle::experimental::SharedStorage>(
-          paddle::memory::Alloc(place, bytes_size), 0),
+          paddle::memory::Alloc(place, bytes_size)),
       std::move(ret_meta));
 
   float* t_ptr = t_dense->mutable_data<float>();

paddle/fluid/eager/tests/task_tests/hook_test.cc

@@ -42,7 +42,7 @@ egr::EagerTensor hook_function(const egr::EagerTensor& t) {
       paddle::framework::product(t_dense->dims()) * SizeOf(t_dense->dtype());
   auto ret_dense = std::make_shared<pten::DenseTensor>(
       pten::make_intrusive<paddle::experimental::SharedStorage>(
-          paddle::memory::Alloc(place, bytes_size), 0),
+          paddle::memory::Alloc(place, bytes_size)),
       std::move(ret_meta));
 
   float* t_ptr = t_dense->mutable_data<float>();

paddle/fluid/framework/custom_operator.cc

@@ -207,14 +207,14 @@ static void RunKernelFunc(const framework::ExecutionContext& ctx,
                 "Tensors.",
                 vec_true_outs.size(), outs.size()));
         for (size_t j = 0; j < vec_true_outs.size(); ++j) {
-          experimental::MovesSharedStorage(
+          experimental::SharesStorage(
               std::dynamic_pointer_cast<pten::DenseTensor>(outs.at(j).impl())
                   .get(),
               vec_true_outs.at(j));
         }
       } else {
         auto* true_out = ctx.Output<Tensor>(out_name);
-        experimental::MovesSharedStorage(
+        experimental::SharesStorage(
             std::dynamic_pointer_cast<pten::DenseTensor>(outs.at(i).impl())
                 .get(),
             true_out);

paddle/fluid/framework/tensor.h

@@ -273,6 +273,7 @@ class Tensor {
 
   const std::shared_ptr<memory::Allocation>& Holder() const { return holder_; }
   size_t offset() const { return offset_; }
+  void set_offset(size_t offset) { offset_ = offset; }
 
   std::shared_ptr<memory::Allocation> MoveMemoryHolder() {
     return std::move(holder_);

paddle/fluid/operators/reshape_op.cc

@@ -456,8 +456,7 @@ class ReshapeKernel {
     // non-inplace need move all result from pt_out to out, inplace need set
     // result dims.
     if (in != out) {
-      paddle::experimental::MovesSharedStorage(pt_out,
-                                               static_cast<Tensor *>(out));
+      paddle::experimental::SharesStorage(pt_out, static_cast<Tensor *>(out));
     } else {
       out->Resize(pt_out->dims());
     }

paddle/pten/api/include/tensor.h

@@ -304,7 +304,7 @@ class PADDLE_API Tensor final {
    *                 The index number begins from begin_idx + 1.
    * @return Tensor
    */
-  Tensor slice(const int64_t begin_idx, const int64_t end_idx) const;
+  Tensor slice(int64_t begin_idx, int64_t end_idx) const;
 
   /**
    * @brief Return the implemention of current Tensor.

paddle/pten/api/lib/tensor.cc

@@ -253,11 +253,11 @@ template PADDLE_API paddle::platform::float16 *
 Tensor::data<paddle::platform::float16>();
 
 // TODO(chenweihang): replace slice impl by API
-Tensor Tensor::slice(const int64_t begin_idx, const int64_t end_idx) const {
+Tensor Tensor::slice(int64_t begin_idx, int64_t end_idx) const {
   if (is_dense_tensor()) {
     return Tensor(std::make_shared<pten::DenseTensor>(
         std::move(pten::CompatibleDenseTensorUtils::Slice(
-            std::dynamic_pointer_cast<pten::DenseTensor>(impl_).get(),
+            *(std::dynamic_pointer_cast<pten::DenseTensor>(impl_).get()),
             begin_idx,
             end_idx))));
   } else {

paddle/pten/api/lib/utils/storage.h

@@ -37,7 +37,6 @@ class ExternalStorage : public pten::Storage {
   void Clear() override {
     data_ = nullptr;
     size_ = 0;
-    offset_ = 0;
   }
 
   size_t size() const noexcept override { return size_; }
@@ -57,13 +56,11 @@ class ExternalStorage : public pten::Storage {
 class SharedStorage : public pten::Storage {
  public:
   explicit SharedStorage(
-      const std::shared_ptr<paddle::memory::Allocation>& allocation,
-      size_t offset)
+      const std::shared_ptr<paddle::memory::Allocation>& allocation)
       : Storage(allocation) {
     CHECK(allocation);
     place_ = allocation->place();
     size_ = allocation->size();
-    offset_ = offset;
   }
 
   // In order to be compatible with the original Tensor design and execution
@@ -84,7 +81,6 @@ class SharedStorage : public pten::Storage {
   void Clear() override {
     data_ = nullptr;
     size_ = 0;
-    offset_ = 0;
   }
 
   size_t size() const noexcept override { return size_; }
@@ -96,12 +92,10 @@ class SharedStorage : public pten::Storage {
   }
 
   // Temporary method: For compatible with fluid Tensor and improve performance
-  void ResetAllocation(std::shared_ptr<paddle::memory::Allocation> allocation,
-                       size_t offset) {
+  void ResetAllocation(std::shared_ptr<paddle::memory::Allocation> allocation) {
     data_ = allocation;
     size_ = allocation->size();
     place_ = allocation->place();
-    offset_ = offset;
   }
 
   // Temporary method: For compatible with fluid Tensor and improve performance