[Refactoring Tensor PR #5] replace storage with pten allocation (Padd…

…lePaddle#39085) * updates callers, test=develop * updates tensor, test=develop * fixes errors, test=develop * remove some dtypes, test=develop * fix errors in the base storage modification, test=develop * fixes a bug, test=develop * fixes the bugs in push the whole, test=develop * updates, test=develop * update * update, test=develop * fixes the mac-py3 CI, test=develop * remove the storage impl, test=develop * updates some codes, test=develop * update, test=develop * updates pten allocation, test=develop
seemingwang · Jan 24, 2022 · a56e16a · a56e16a
1 parent 95b081e
commit a56e16a
Show file tree

Hide file tree

Showing 80 changed files with 490 additions and 684 deletions.
diff --git a/paddle/fluid/eager/accumulation/gradient_accumulation.cc b/paddle/fluid/eager/accumulation/gradient_accumulation.cc
@@ -216,8 +216,9 @@ void TensorAdd(const egr::EagerTensor& src, egr::EagerTensor* dst) {
 
 #define PADDLE_TENSOR_ADD(cpp_type)                                          \
   if (data_type == paddle::framework::DataTypeTrait<cpp_type>::DataType()) { \
-    TensorAddFunctor<cpp_type> func(numel, src_tensor->data<cpp_type>(),     \
-                                    dst_tensor->mutable_data<cpp_type>());   \
+    TensorAddFunctor<cpp_type> func(                                         \
+        numel, src_tensor->data<cpp_type>(),                                 \
+        dst_tensor->mutable_data<cpp_type>(place));                          \
     paddle::platform::VisitPlace(place, func);                               \
     return;                                                                  \
   }

diff --git a/paddle/fluid/eager/tests/data_structure_tests/accumulation_node_test.cc b/paddle/fluid/eager/tests/data_structure_tests/accumulation_node_test.cc
@@ -36,7 +36,8 @@ TEST(AccumulationNode, EagerTensor) {
           paddle::platform::CPUPlace())
           .get(),
       meta);
-  dt0->mutable_data<paddle::platform::float16>()[0] = 10.0;
+  dt0->mutable_data<paddle::platform::float16>(
+      paddle::platform::CPUPlace())[0] = 10.0;
   EagerTensor et0 = EagerTensor(dt0);
 
   std::shared_ptr<pten::DenseTensor> dt1 = std::make_shared<pten::DenseTensor>(
@@ -45,7 +46,8 @@ TEST(AccumulationNode, EagerTensor) {
           .get(),
       meta);
 
-  dt1->mutable_data<paddle::platform::float16>()[0] = 20.0;
+  dt1->mutable_data<paddle::platform::float16>(
+      paddle::platform::CPUPlace())[0] = 20.0;
   EagerTensor et1 = EagerTensor(dt1);
 
   std::shared_ptr<pten::DenseTensor> grad_dt =

diff --git a/paddle/fluid/eager/tests/data_structure_tests/autograd_meta_test.cc b/paddle/fluid/eager/tests/data_structure_tests/autograd_meta_test.cc
@@ -46,7 +46,7 @@ TEST(AutogradMeta, MemberFunction) {
           paddle::platform::CPUPlace())
           .get(),
       meta);
-  auto* dt_ptr = dt->mutable_data<float>();
+  auto* dt_ptr = dt->mutable_data<float>(paddle::platform::CPUPlace());
   dt_ptr[0] = 5.0f;
   dt_ptr[1] = 10.0f;
   grad_t->set_impl(dt);

diff --git a/paddle/fluid/eager/tests/data_structure_tests/eager_tensor_test.cc b/paddle/fluid/eager/tests/data_structure_tests/eager_tensor_test.cc
@@ -40,7 +40,7 @@ TEST(EagerTensor, Constructor) {
           paddle::platform::CPUPlace())
           .get(),
       meta);
-  auto* dt_ptr = dt->mutable_data<float>();
+  auto* dt_ptr = dt->mutable_data<float>(paddle::platform::CPUPlace());
   dt_ptr[0] = 5.0f;
   dt_ptr[1] = 10.0f;
   egr::EagerTensor et3 = egr::EagerTensor(dt);
@@ -70,7 +70,7 @@ TEST(EagerTensor, MemberFunction) {
           paddle::platform::CPUPlace())
           .get(),
       meta);
-  auto* dt_ptr = dt->mutable_data<float>();
+  auto* dt_ptr = dt->mutable_data<float>(paddle::platform::CPUPlace());
   dt_ptr[0] = 5.0f;
   dt_ptr[1] = 10.0f;
   VLOG(6) << "Make Dense Tensor";

diff --git a/paddle/fluid/eager/tests/data_structure_tests/grad_node_info_test.cc b/paddle/fluid/eager/tests/data_structure_tests/grad_node_info_test.cc
@@ -45,7 +45,7 @@ TEST(GradNodeInfo, GradNodeBase) {
           paddle::platform::CPUPlace())
           .get(),
       meta);
-  auto* dt_ptr = dt->mutable_data<float>();
+  auto* dt_ptr = dt->mutable_data<float>(paddle::platform::CPUPlace());
   dt_ptr[0] = 5.0f;
   egr::EagerTensor et1(dt);
   grads = {{et1}};
@@ -102,7 +102,7 @@ TEST(GradNodeInfo, GradNodeBase) {
             paddle::platform::CPUPlace())
             .get(),
         meta);
-    auto* dt_ptr = dt->mutable_data<float>();
+    auto* dt_ptr = dt->mutable_data<float>(paddle::platform::CPUPlace());
     dt_ptr[0] = 6.0f;
     auto* et_ptr =
         std::dynamic_pointer_cast<pten::DenseTensor>(et.impl())->data<float>();
@@ -121,8 +121,8 @@ TEST(GradNodeInfo, GradNodeBase) {
 
   VLOG(6) << "Test Reduce Hook";
   auto reduce_hook = [&](void) -> void {
-    auto* et_ptr = std::dynamic_pointer_cast<pten::DenseTensor>(et1.impl())
-                       ->mutable_data<float>();
+    auto* et_ptr =
+        std::dynamic_pointer_cast<pten::DenseTensor>(et1.impl())->data<float>();
     et_ptr[0] = 100.0;
     VLOG(6) << "Running Reduce Hook";
   };

diff --git a/paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h b/paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h
@@ -41,7 +41,7 @@ class GradTestNode : public egr::GradNodeBase {
             paddle::platform::CPUPlace())
             .get(),
         meta);
-    auto* dt_ptr = dt->mutable_data<float>();
+    auto* dt_ptr = dt->mutable_data<float>(paddle::platform::CPUPlace());
     dt_ptr[0] = 6.0f;
     egr::EagerTensor et1(dt);
     std::vector<std::vector<egr::EagerTensor>> res = {{et1}};

diff --git a/paddle/fluid/eager/tests/data_structure_tests/grad_tensor_holder_test.cc b/paddle/fluid/eager/tests/data_structure_tests/grad_tensor_holder_test.cc
@@ -57,15 +57,15 @@ TEST(GradTensorHolder, Interfaces) {
           paddle::platform::CPUPlace())
           .get(),
       meta);
-  dt0->mutable_data<float>()[0] = 10.0;
+  dt0->mutable_data<float>(paddle::platform::CPUPlace())[0] = 10.0;
   EagerTensor et0 = EagerTensor(dt0);
 
   std::shared_ptr<pten::DenseTensor> dt1 = std::make_shared<pten::DenseTensor>(
       std::make_unique<paddle::experimental::DefaultAllocator>(
           paddle::platform::CPUPlace())
           .get(),
       meta);
-  dt1->mutable_data<float>()[0] = 20.0;
+  dt1->mutable_data<float>(paddle::platform::CPUPlace())[0] = 20.0;
   EagerTensor et1 = EagerTensor(dt1);
 
   // Constructor empty GradTensorHolder

diff --git a/paddle/fluid/eager/tests/data_structure_tests/tensor_wrapper_test.cc b/paddle/fluid/eager/tests/data_structure_tests/tensor_wrapper_test.cc
@@ -29,7 +29,7 @@ TEST(TensorWrapper, Basic) {
           paddle::platform::CPUPlace())
           .get(),
       meta);
-  auto* dt_ptr = dt->mutable_data<float>();
+  auto* dt_ptr = dt->mutable_data<float>(paddle::platform::CPUPlace());
   dt_ptr[0] = 5.0f;
   dt_ptr[1] = 10.0f;
   et1.set_impl(dt);
@@ -56,7 +56,7 @@ TEST(TensorWrapper, Basic) {
           paddle::platform::CPUPlace())
           .get(),
       meta2);
-  auto* dt_ptr2 = dt->mutable_data<float>();
+  auto* dt_ptr2 = dt->mutable_data<float>(paddle::platform::CPUPlace());
   dt_ptr2[0] = 6.0f;
   dt_ptr2[1] = 11.0f;
   et2.set_impl(dt2);

diff --git a/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc b/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc
@@ -35,15 +35,15 @@ TEST(EagerUtils, AutoGradMeta) {
           paddle::platform::CPUPlace())
           .get(),
       meta);
-  dt0->mutable_data<float>()[0] = 10.0;
+  dt0->mutable_data<float>(paddle::platform::CPUPlace())[0] = 10.0;
   EagerTensor et0 = EagerTensor(dt0);
 
   std::shared_ptr<pten::DenseTensor> dt1 = std::make_shared<pten::DenseTensor>(
       std::make_unique<paddle::experimental::DefaultAllocator>(
           paddle::platform::CPUPlace())
           .get(),
       meta);
-  dt1->mutable_data<float>()[0] = 20.0;
+  dt1->mutable_data<float>(paddle::platform::CPUPlace())[0] = 20.0;
   EagerTensor et1 = EagerTensor(dt1);
 
   std::vector<EagerTensor> ets = {et0, et1};
@@ -112,7 +112,7 @@ egr::EagerTensor CreateTestCPUTensor(T val,
           paddle::platform::CPUPlace())
           .get(),
       meta);
-  auto* dt_ptr = dt->mutable_data<T>();
+  auto* dt_ptr = dt->mutable_data<T>(paddle::platform::CPUPlace());
   for (int64_t i = 0; i < dt->numel(); i++) {
     dt_ptr[i] = val;
   }

diff --git a/paddle/fluid/eager/tests/task_tests/fwd_bwd_joint_test.cc b/paddle/fluid/eager/tests/task_tests/fwd_bwd_joint_test.cc
@@ -44,8 +44,8 @@ egr::EagerTensor hook_function(const egr::EagerTensor& t) {
           paddle::memory::Alloc(place, bytes_size)),
       std::move(ret_meta));
 
-  float* t_ptr = t_dense->mutable_data<float>();
-  float* ret_ptr = ret_dense->mutable_data<float>();
+  float* t_ptr = t_dense->mutable_data<float>(place);
+  float* ret_ptr = ret_dense->mutable_data<float>(place);
   for (int i = 0; i < ret_dense->numel(); i++) {
     ret_ptr[i] = t_ptr[i] + 5.0;
   }
@@ -184,7 +184,7 @@ TEST(FwdBwdJoint, BranchedNodes) {
   // Examine Forward Output 2
   {
     auto dense_out = std::dynamic_pointer_cast<pten::DenseTensor>(out2.impl());
-    float* ptr = dense_out->mutable_data<float>();
+    float* ptr = dense_out->mutable_data<float>(paddle::platform::CPUPlace());
     for (int i = 0; i < 20; i++) {
       PADDLE_ENFORCE(ptr[i] == 150.0,
                      paddle::platform::errors::Fatal(

diff --git a/paddle/fluid/eager/tests/task_tests/hook_test.cc b/paddle/fluid/eager/tests/task_tests/hook_test.cc
@@ -45,8 +45,8 @@ egr::EagerTensor hook_function(const egr::EagerTensor& t) {
           paddle::memory::Alloc(place, bytes_size)),
       std::move(ret_meta));
 
-  float* t_ptr = t_dense->mutable_data<float>();
-  float* ret_ptr = ret_dense->mutable_data<float>();
+  float* t_ptr = t_dense->mutable_data<float>(place);
+  float* ret_ptr = ret_dense->mutable_data<float>(place);
   for (int i = 0; i < ret_dense->numel(); i++) {
     ret_ptr[i] = t_ptr[i] + 3.0;
   }

diff --git a/paddle/fluid/eager/tests/test_utils.h b/paddle/fluid/eager/tests/test_utils.h
@@ -34,7 +34,7 @@ bool CompareGradTensorWithValue(const egr::EagerTensor& target, T value) {
   egr::AutogradMeta* meta = egr::EagerUtils::unsafe_autograd_meta(target);
   auto grad_dense =
       std::dynamic_pointer_cast<pten::DenseTensor>(meta->Grad().impl());
-  T* ptr = grad_dense->mutable_data<T>();
+  T* ptr = grad_dense->data<T>();
 
   std::vector<T> host_data(grad_dense->numel());
   if (paddle::platform::is_gpu_place(grad_dense->place())) {
@@ -67,7 +67,7 @@ template <typename T>
 bool CompareTensorWithValue(const egr::EagerTensor& target, T value) {
   // TODO(jiabin): Support Selected Rows later
   auto dense_t = std::dynamic_pointer_cast<pten::DenseTensor>(target.impl());
-  T* ptr = dense_t->mutable_data<T>();
+  T* ptr = dense_t->data<T>();
 
   std::vector<T> host_data(dense_t->numel());
   if (paddle::platform::is_gpu_place(dense_t->place())) {

diff --git a/paddle/fluid/pybind/eager.cc b/paddle/fluid/pybind/eager.cc
@@ -80,6 +80,7 @@ void EmptyEagerTensorInitializer(
         std::make_shared<pten::DenseTensor>(
             pten::make_intrusive<paddle::experimental::SharedStorage>(place),
             pten::DenseTensorMeta(pten::TransToPtenDataType(dtype), ddims));
+    dense_tensor->mutable_data(place);
     self->eager_tensor.set_impl(dense_tensor);
   } else {
     PADDLE_THROW(platform::errors::InvalidArgument(

diff --git a/paddle/pten/api/ext/dispatch.h b/paddle/pten/api/ext/dispatch.h
@@ -272,16 +272,10 @@ namespace paddle {
           NAME, ::pten::DataType::UINT8, uint8_t, __VA_ARGS__)                \
       PD_PRIVATE_CASE_TYPE(                                                   \
           NAME, ::pten::DataType::INT16, int16_t, __VA_ARGS__)                \
-      PD_PRIVATE_CASE_TYPE(                                                   \
-          NAME, ::pten::DataType::UINT16, uint16_t, __VA_ARGS__)              \
       PD_PRIVATE_CASE_TYPE(                                                   \
           NAME, ::pten::DataType::INT32, int32_t, __VA_ARGS__)                \
-      PD_PRIVATE_CASE_TYPE(                                                   \
-          NAME, ::pten::DataType::UINT32, uint32_t, __VA_ARGS__)              \
       PD_PRIVATE_CASE_TYPE(                                                   \
           NAME, ::pten::DataType::INT64, int64_t, __VA_ARGS__)                \
-      PD_PRIVATE_CASE_TYPE(                                                   \
-          NAME, ::pten::DataType::UINT64, uint64_t, __VA_ARGS__)              \
       PD_PRIVATE_CASE_TYPE(NAME,                                              \
                            ::pten::DataType::BFLOAT16,                        \
                            paddle::experimental::bfloat16,                    \

diff --git a/paddle/pten/api/lib/tensor.cc b/paddle/pten/api/lib/tensor.cc
@@ -149,8 +149,8 @@ bool Tensor::is_cuda() const {
 template <typename T>
 T *Tensor::mutable_data() {
   if (is_dense_tensor()) {
-    return std::dynamic_pointer_cast<pten::DenseTensor>(impl_)
-        ->mutable_data<T>();
+    return std::dynamic_pointer_cast<pten::DenseTensor>(impl_)->mutable_data<T>(
+        ConvertExtPlaceToInnerPlace(place()));
   }
   return nullptr;
 }
@@ -173,12 +173,18 @@ Tensor::mutable_data<paddle::platform::float16>();
 template <typename T>
 T *Tensor::mutable_data(const PlaceType &place) {
   auto inner_place = ConvertExtPlaceToInnerPlace(place);
-  PADDLE_ENFORCE_EQ(
-      platform::is_same_place(inner_place, impl_->place()),
-      true,
-      platform::errors::Unimplemented("Modification of tensor place through "
-                                      "mutable_data is not supported now"));
-  return mutable_data<T>();
+  if (impl_->initialized()) {
+    PADDLE_ENFORCE_EQ(
+        platform::is_same_place(inner_place, impl_->place()),
+        true,
+        platform::errors::Unimplemented("Modification of tensor place through "
+                                        "mutable_data is not supported now"));
+  }
+  if (is_dense_tensor()) {
+    return std::dynamic_pointer_cast<pten::DenseTensor>(impl_)->mutable_data<T>(
+        inner_place);
+  }
+  return nullptr;
 }
 
 template PADDLE_API float *Tensor::mutable_data<float>(const PlaceType &place);
@@ -205,7 +211,8 @@ Tensor::mutable_data<paddle::platform::float16>(const PlaceType &place);
 template <typename T>
 const T *Tensor::data() const {
   if (is_dense_tensor()) {
-    return std::dynamic_pointer_cast<pten::DenseTensor>(impl_)->data<T>();
+    return std::dynamic_pointer_cast<pten::DenseTensor>(impl_)->mutable_data<T>(
+        ConvertExtPlaceToInnerPlace(place()));
   }
   return nullptr;
 }
@@ -217,7 +224,6 @@ template PADDLE_API const int32_t *Tensor::data<int32_t>() const;
 template PADDLE_API const uint8_t *Tensor::data<uint8_t>() const;
 template PADDLE_API const int8_t *Tensor::data<int8_t>() const;
 template PADDLE_API const int16_t *Tensor::data<int16_t>() const;
-template PADDLE_API const uint16_t *Tensor::data<uint16_t>() const;
 template PADDLE_API const bool *Tensor::data<bool>() const;
 template PADDLE_API const paddle::platform::complex<float>
     *Tensor::data<paddle::platform::complex<float>>() const;

diff --git a/paddle/pten/api/lib/utils.cc b/paddle/pten/api/lib/utils.cc
@@ -65,6 +65,7 @@ PADDLE_API Tensor copy_to(const Tensor& x, Backend backend, bool blocking) {
       pten::make_intrusive<paddle::experimental::SharedStorage>(
           pten::TransToFluidPlace(backend)),
       std::move(out_meta));
+  dense_out->mutable_data(pten::TransToFluidPlace(backend));
   kernel_context.EmplaceBackOutput(dense_out.get());
   Tensor out;
   out.set_impl(dense_out);

diff --git a/paddle/pten/api/lib/utils/storage.h b/paddle/pten/api/lib/utils/storage.h
@@ -39,6 +39,18 @@ class ExternalStorage : public pten::Storage {
     size_ = 0;
   }
 
+  void set_data_shared(
+      const std::shared_ptr<paddle::memory::Allocation>& holder) override {
+    CHECK(holder);
+    data_ = holder;
+    size_ = holder->size();
+  }
+
+  std::shared_ptr<paddle::memory::Allocation>&& move_data_shared() override {
+    size_ = 0;
+    return std::move(data_);
+  }
+
   size_t size() const noexcept override { return size_; }
   const paddle::platform::Place& place() const override {
     PADDLE_ENFORCE_NOT_NULL(
@@ -92,6 +104,12 @@ class SharedStorage : public pten::Storage {
     }
   }
 
+  std::shared_ptr<paddle::memory::Allocation>&& move_data_shared() override {
+    size_ = 0;
+    place_ = Place();
+    return std::move(data_);
+  }
+
   size_t size() const noexcept override {
     return data_ ? data_->size() : size_;
   }

diff --git a/paddle/pten/common/scalar.h b/paddle/pten/common/scalar.h
@@ -133,9 +133,6 @@ class ScalarBase {
       case DataType::INT8:
         data_.i8 = tensor.template data<int8_t>()[0];
         break;
-      case DataType::UINT16:
-        data_.ui16 = tensor.template data<uint16_t>()[0];
-        break;
       case DataType::UINT8:
         data_.ui8 = tensor.template data<uint8_t>()[0];
         break;