Revert D13596031: Improve c2-aten tensor interop and add proper testing

Differential Revision:
D13596031

Original commit changeset: d20b601e06ba

fbshipit-source-id: dc371697f14b3893a9164380a39e7a49d8d68ecf

9 files changed, 54 insertions, 400 deletions

diff --git a/aten/src/ATen/core/LegacyTypeDispatch.h b/aten/src/ATen/core/LegacyTypeDispatch.h
index 66bfcd65d5..9c9d44e84f 100644
--- a/aten/src/ATen/core/LegacyTypeDispatch.h
+++ b/aten/src/ATen/core/LegacyTypeDispatch.h
@@ -176,20 +176,9 @@ struct CAFFE2_API AutoNonVariableTypeMode {
 inline Type& legacyTensorType(const TensorImpl& tensor) {
   // NB: It's valid to use getTypeRaw here, because the TensorImpl
   // could not have been created without initializing the Type first.
-  // NB: This is not actually true via the Caffe2 codepath! But we call
-  // initializeLegacyTypeDispatchFor in the right place.
-  return *globalLegacyTypeDispatch().getTypeRaw(
-      tensorTypeIdToBackend(tensor.type_id()),
-      typeMetaToScalarType(tensor.dtype()),
-      tensor.is_variable() && !at::NonVariableTypeMode::is_enabled());
-}
-
-inline void initializeLegacyTypeDispatchFor(const TensorImpl& tensor) {
-  // getType calls the right initialization
-  globalLegacyTypeDispatch().getType(
-      tensorTypeIdToBackend(tensor.type_id()),
-      typeMetaToScalarType(tensor.dtype()),
-      tensor.is_variable() && !at::NonVariableTypeMode::is_enabled());
+  // TODO: This is not actually true via the Caffe2 codepath!  Make
+  // it so.
+  return *globalLegacyTypeDispatch().getTypeRaw(tensorTypeIdToBackend(tensor.type_id()), typeMetaToScalarType(tensor.dtype()), tensor.is_variable() && !at::NonVariableTypeMode::is_enabled());
 }
 
 } // namespace at
diff --git a/aten/src/ATen/core/Tensor.cpp b/aten/src/ATen/core/Tensor.cpp
index a6489c0a4e..924688d40b 100644
--- a/aten/src/ATen/core/Tensor.cpp
+++ b/aten/src/ATen/core/Tensor.cpp
@@ -6,27 +6,6 @@
 
 namespace at {
 
-void Tensor::enforce_invariants() {
-  if (impl_.get() == nullptr) {
-    throw std::runtime_error("TensorImpl with nullptr is not supported");
-  }
-  // Following line throws if the method is not a POD data type or is not
-  // supported by ATen
-  scalar_type();
-  if (defined()) {
-    AT_ASSERTM(
-        impl_->dtype_initialized(),
-        "Partially-initialized tensor not supported by at::Tensor");
-    AT_ASSERTM(
-        impl_->storage_initialized(),
-        "Partially-initialized tensor not supported by at::Tensor");
-    // Ensure LegacyTypeDispatch is initialized. In ATen it's done in tensor
-    // factory functions, but when we get a tensor from Caffe2 we might bypass
-    // those factory functions.
-    initializeLegacyTypeDispatchFor(*impl_);
-  }
-}
-
 void Tensor::print() const {
   if (defined()) {
     std::cerr << "[" << type().toString() << " " << sizes() << "]" << std::endl;
diff --git a/aten/src/ATen/core/Tensor.h b/aten/src/ATen/core/Tensor.h
index ba4ccb4393..6abf5fcbe9 100644
--- a/aten/src/ATen/core/Tensor.h
+++ b/aten/src/ATen/core/Tensor.h
@@ -46,34 +46,20 @@ using TensorList = ArrayRef<Tensor>;
 // Note that Tensor can also be NULL, i.e. it is not associated with any underlying TensorImpl, and
 // special care must be taken to handle this.
 class CAFFE2_API Tensor {
- public:
+public:
   Tensor(){};
-  // This constructor should not be used by end users and is an implementation
-  // detail invoked by autogenerated code.
-  explicit Tensor(
-      c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> tensor_impl)
+  Tensor(c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> tensor_impl)
       : impl_(std::move(tensor_impl)) {
     if (impl_.get() == nullptr) {
-      throw std::runtime_error("TensorImpl with nullptr is not supported");
+      throw std::runtime_error("TensorBaseImpl with nullptr not supported");
     }
   }
+
   Tensor(const Tensor&) = default;
   Tensor(Tensor&&) = default;
 
-
- public:
-  // Creates a new wrapper from TensorImpl. Intentionally a free method because
-  // it should be used with care. Checks necessary invariants
-  static Tensor wrap_tensor_impl(
-      c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> tensor_impl) {
-    Tensor r(std::move(tensor_impl));
-    r.enforce_invariants();
-    return r;
-  }
-
-  explicit Tensor(C10Tensor tensor) : impl_(std::move(tensor).impl()) {
-    enforce_invariants();
-  }
+  explicit Tensor(C10Tensor tensor)
+      : impl_(std::move(tensor).impl()) {}
 
   explicit operator C10Tensor() const & {
     return C10Tensor(impl_);
@@ -722,7 +708,6 @@ class CAFFE2_API Tensor {
   friend struct WeakTensor;
 
 protected:
-  void enforce_invariants();
   c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> impl_;
 };
 
diff --git a/aten/src/ATen/templates/Tensor.h b/aten/src/ATen/templates/Tensor.h
index ce69ef4bd5..fe0cd0c157 100644
--- a/aten/src/ATen/templates/Tensor.h
+++ b/aten/src/ATen/templates/Tensor.h
@@ -46,34 +46,20 @@ using TensorList = ArrayRef<Tensor>;
 // Note that Tensor can also be NULL, i.e. it is not associated with any underlying TensorImpl, and
 // special care must be taken to handle this.
 class CAFFE2_API Tensor {
- public:
+public:
   Tensor(){};
-  // This constructor should not be used by end users and is an implementation
-  // detail invoked by autogenerated code.
-  explicit Tensor(
-      c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> tensor_impl)
+  Tensor(c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> tensor_impl)
       : impl_(std::move(tensor_impl)) {
     if (impl_.get() == nullptr) {
-      throw std::runtime_error("TensorImpl with nullptr is not supported");
+      throw std::runtime_error("TensorBaseImpl with nullptr not supported");
     }
   }
+
   Tensor(const Tensor&) = default;
   Tensor(Tensor&&) = default;
 
-
- public:
-  // Creates a new wrapper from TensorImpl. Intentionally a free method because
-  // it should be used with care. Checks necessary invariants
-  static Tensor wrap_tensor_impl(
-      c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> tensor_impl) {
-    Tensor r(std::move(tensor_impl));
-    r.enforce_invariants();
-    return r;
-  }
-
-  explicit Tensor(C10Tensor tensor) : impl_(std::move(tensor).impl()) {
-    enforce_invariants();
-  }
+  explicit Tensor(C10Tensor tensor)
+      : impl_(std::move(tensor).impl()) {}
 
   explicit operator C10Tensor() const & {
     return C10Tensor(impl_);
@@ -327,7 +313,6 @@ class CAFFE2_API Tensor {
   friend struct WeakTensor;
 
 protected:
-  void enforce_invariants();
   c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> impl_;
 };
 
diff --git a/aten/src/ATen/test/CMakeLists.txt b/aten/src/ATen/test/CMakeLists.txt
index 1d2457154b..d3f0d0134b 100644
--- a/aten/src/ATen/test/CMakeLists.txt
+++ b/aten/src/ATen/test/CMakeLists.txt
@@ -29,8 +29,7 @@ list(APPEND ATen_CUDA_TEST_SRCS
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_stream_test.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_half_test.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_optional_test.cu
-  ${CMAKE_CURRENT_SOURCE_DIR}/cuda_packedtensoraccessor_test.cu
-  ${CMAKE_CURRENT_SOURCE_DIR}/cuda_tensor_interop_test.cpp)
+  ${CMAKE_CURRENT_SOURCE_DIR}/cuda_packedtensoraccessor_test.cu)
 if (CUDNN_FOUND)
   list(APPEND ATen_CUDA_TEST_SRCS
     ${CMAKE_CURRENT_SOURCE_DIR}/cuda_cudnn_test.cpp)
diff --git a/aten/src/ATen/test/cuda_tensor_interop_test.cpp b/aten/src/ATen/test/cuda_tensor_interop_test.cpp
deleted file mode 100644
index 9877fd2e03..0000000000
--- a/aten/src/ATen/test/cuda_tensor_interop_test.cpp
+++ /dev/null
@@ -1,164 +0,0 @@
-#include "gtest/gtest.h"
-
-#include "ATen/ATen.h"
-#include <ATen/cuda/CUDAContext.h>
-#include <caffe2/core/init.h>
-#include <caffe2/core/operator.h>
-#include <caffe2/core/context_gpu.h>
-#include <caffe2/utils/math.h>
-
-// dumbest possible copies
-template<typename T>
-T cuda_get(T* addr) {
-  T result;
-  CUDA_ENFORCE(cudaMemcpy(&result, addr, sizeof(T), cudaMemcpyDefault));
-  return result;
-}
-
-template<typename T>
-void cuda_set(T* addr, T value) {
-  CUDA_ENFORCE(cudaMemcpy(addr, &value, sizeof(T), cudaMemcpyDefault));
-}
-
-TEST(CUDACaffe2ToPytorch, SimpleLegacy) {
-  if (!at::cuda::is_available()) return;
-  caffe2::Tensor c2_tensor(caffe2::CUDA);
-  c2_tensor.Resize(4, 4);
-  auto data = c2_tensor.mutable_data<int64_t>();
-  {
-    caffe2::CUDAContext context;
-    caffe2::math::Set<int64_t>(16, 777, data, &context);
-  }
-  at::Tensor at_tensor(c2_tensor);
-  ASSERT_TRUE(&at_tensor.type() != nullptr);
-  ASSERT_TRUE(at_tensor.is_cuda());
-
-  auto at_cpu = at_tensor.cpu();
-  auto it = at_cpu.data<int64_t>();
-  for (int64_t i = 0; i < 16; i++) {
-    ASSERT_EQ(it[i], 777);
-  }
-}
-
-TEST(CUDACaffe2ToPytorch, Simple) {
-  if (!at::cuda::is_available()) return;
-  caffe2::Tensor c2_tensor =
-      caffe2::empty({4, 4}, at::dtype<int64_t>().device(caffe2::CUDA));
-  auto data = c2_tensor.mutable_data<int64_t>();
-  {
-    caffe2::CUDAContext context;
-    caffe2::math::Set<int64_t>(16, 777, data, &context);
-  }
-  at::Tensor at_tensor(c2_tensor);
-  ASSERT_TRUE(&at_tensor.type() != nullptr);
-  ASSERT_TRUE(at_tensor.is_cuda());
-
-  auto at_cpu = at_tensor.cpu();
-  auto it = at_cpu.data<int64_t>();
-  for (int64_t i = 0; i < 16; i++) {
-    ASSERT_EQ(it[i], 777);
-  }
-}
-
-TEST(CUDACaffe2ToPytorch, Op) {
-  if (!at::cuda::is_available()) return;
-  caffe2::Tensor c2_tensor =
-      caffe2::empty({3, 3}, at::dtype<int64_t>().device(caffe2::CUDA));
-  auto data = c2_tensor.mutable_data<int64_t>();
-  {
-    caffe2::CUDAContext context;
-    caffe2::math::Set<int64_t>(9, 111, data, &context);
-  }
-  at::Tensor at_tensor(c2_tensor);
-  ASSERT_TRUE(at_tensor.is_cuda());
-
-  ASSERT_EQ(at::sum(at_tensor).item<int64_t>(), 999);
-}
-
-TEST(CUDAPytorchToCaffe2, Op) {
-  if (!at::cuda::is_available()) return;
-  caffe2::Workspace workspace;
-  caffe2::NetDef net;
-
-  auto at_tensor_a = at::ones({5, 5}, at::dtype(at::kFloat).device(at::kCUDA));
-  auto at_tensor_b = at::ones({5, 5}, at::dtype(at::kFloat).device(at::kCUDA));
-  auto at_tensor_c = at::ones({5, 5}, at::dtype(at::kFloat).device(at::kCUDA));
-
-  auto* c2_tensor_a = BlobSetTensor(workspace.CreateBlob("a"), caffe2::Tensor(at_tensor_a));
-  auto* c2_tensor_b = BlobSetTensor(workspace.CreateBlob("b"), caffe2::Tensor(at_tensor_b));
-
-  // Test Alias
-  {
-    caffe2::Tensor c2_tensor_from_aten(at_tensor_c);
-    BlobSetTensor(workspace.CreateBlob("c"), c2_tensor_from_aten.Alias());
-  }
-
-  {
-    auto op = net.add_op();
-    op->set_type("Sum");
-    op->add_input("a");
-    op->add_input("b");
-    op->add_input("c");
-    op->add_output("d");
-    op->mutable_device_option()->set_device_type(caffe2::PROTO_CUDA);
-  }
-
-  workspace.RunNetOnce(net);
-
-  const auto& result = workspace.GetBlob("d")->Get<caffe2::Tensor>();
-  ASSERT_EQ(result.GetDeviceType(), caffe2::CUDA);
-
-  auto data = result.data<float>();
-  for (int64_t i = 0; i < 25; i++) {
-    ASSERT_EQ(cuda_get(data + i), 3.0);
-  }
-  at::Tensor at_result(result);
-  ASSERT_TRUE(at_result.is_cuda());
-  ASSERT_EQ(at::sum(at_result).item<float>(), 75);
-}
-
-TEST(CUDAPytorchToCaffe2, SharedStorageWrite) {
-  if (!at::cuda::is_available()) return;
-  auto at_tensor_a = at::ones({5, 5}, at::dtype(at::kFloat).device(at::kCUDA));
-  auto at_tensor_b = at_tensor_a.view({25});
-
-  caffe2::Tensor c2_tensor_a(at_tensor_a);
-  caffe2::Tensor c2_tensor_b(at_tensor_b);
-
-  // change is visible everywhere
-  cuda_set<float>(c2_tensor_a.mutable_data<float>() + 1, 123);
-  ASSERT_EQ(cuda_get(c2_tensor_b.mutable_data<float>() + 1), 123);
-  ASSERT_EQ(at_tensor_a[0][1].item().to<float>(), 123);
-  ASSERT_EQ(at_tensor_b[1].item().to<float>(), 123);
-}
-
-TEST(CUDAPytorchToCaffe2, MutualResizes) {
-  if (!at::cuda::is_available()) return;
-  auto at_tensor = at::ones({5, 5}, at::dtype(at::kFloat).device(at::kCUDA));
-
-  caffe2::Tensor c2_tensor(at_tensor);
-
-  // change is visible
-  cuda_set<float>(c2_tensor.mutable_data<float>(), 123);
-  ASSERT_EQ(at_tensor[0][0].item().to<float>(), 123);
-
-  // resize PT tensor in smaller direction - storage is preserved
-  at_tensor.resize_({4, 4});
-  cuda_set<float>(c2_tensor.mutable_data<float>() + 1, 234);
-  ASSERT_EQ(at_tensor[0][1].item().to<float>(), 234);
-
-  // resize PT tensor in larger direction - storage is preserved
-  at_tensor.resize_({6, 6});
-  cuda_set<float>(c2_tensor.mutable_data<float>() + 2, 345);
-  ASSERT_EQ(at_tensor[0][2].item().to<float>(), 345);
-  ASSERT_EQ(c2_tensor.sizes()[0], 6);
-  ASSERT_EQ(c2_tensor.sizes()[1], 6);
-
-  // resize Caffe2 tensor - semantics are to NOT preserve the data, but the
-  // TensorImpl is still shared
-  c2_tensor.Resize(7, 7);
-  cuda_set<float>(c2_tensor.mutable_data<float>() + 3, 456);
-  ASSERT_EQ(at_tensor[0][3].item().to<float>(), 456);
-  ASSERT_EQ(at_tensor.sizes()[0], 7);
-  ASSERT_EQ(at_tensor.sizes()[1], 7);
-}
diff --git a/aten/src/ATen/test/tensor_interop_test.cpp b/aten/src/ATen/test/tensor_interop_test.cpp
index 6016142db1..f926312b5f 100644
--- a/aten/src/ATen/test/tensor_interop_test.cpp
+++ b/aten/src/ATen/test/tensor_interop_test.cpp
@@ -4,83 +4,58 @@
 #include <caffe2/core/init.h>
 #include <caffe2/core/operator.h>
 
-TEST(Caffe2ToPytorch, SimpleLegacy) {
+TEST(TestTensorInterop, Caffe2ToPytorchSimpleLegacy) {
   caffe2::Tensor c2_tensor(caffe2::CPU);
   c2_tensor.Resize(4, 4);
   auto data = c2_tensor.mutable_data<int64_t>();
   for (int64_t i = 0; i < 16; i++) {
     data[i] = i;
   }
-  at::Tensor at_tensor(c2_tensor);
-  ASSERT_TRUE(&at_tensor.type() != nullptr);
 
-  auto it = at_tensor.data<int64_t>();
+  // TODO: find out why calling data on tensor doesn't work
+  at::Tensor at_tensor(c2_tensor.getIntrusivePtr());
+  at::TensorImpl* impl = at_tensor.unsafeGetTensorImpl();
+
+  auto it = impl->data<int64_t>();
   for (int64_t i = 0; i < 16; i++) {
     ASSERT_EQ(it[i], i);
   }
 }
 
-TEST(Caffe2ToPytorch, Simple) {
+TEST(TestTensorInterop, Caffe2ToPytorchSimple) {
   caffe2::Tensor c2_tensor = caffe2::empty({4, 4}, at::kLong);
   auto data = c2_tensor.mutable_data<int64_t>();
   for (int64_t i = 0; i < 16; i++) {
     data[i] = i;
   }
-  at::Tensor at_tensor(c2_tensor);
-  ASSERT_TRUE(&at_tensor.type() != nullptr);
+  at::Tensor at_tensor(c2_tensor.getIntrusivePtr());
+  at::TensorImpl* impl = at_tensor.unsafeGetTensorImpl();
 
-  auto it = at_tensor.data<int64_t>();
+  auto it = impl->data<int64_t>();
   for (int64_t i = 0; i < 16; i++) {
     ASSERT_EQ(it[i], i);
   }
 }
 
-TEST(Caffe2ToPytorch, Op) {
+TEST(TestTensorInterop, Caffe2ToPytorchOp) {
   caffe2::Tensor c2_tensor(caffe2::CPU);
   c2_tensor.Resize(3, 3);
   auto data = c2_tensor.mutable_data<int64_t>();
   for (int64_t i = 0; i < 9; i++) {
     data[i] = i;
   }
-  at::Tensor at_tensor(c2_tensor);
+  at::Tensor at_tensor(c2_tensor.getIntrusivePtr());
 
   ASSERT_EQ(at::sum(at_tensor).item<int64_t>(), 36);
 }
 
-// Caffe2 doesn't actually have another always-on backend that is not CPU or GPU
-// TEST(Caffe2ToPytorch, UnsupportedDevice) {
-//   caffe2::Tensor c2_tensor(caffe2::OPENGL);
-//   c2_tensor.Resize(4, 4);
-//   c2_tensor.mutable_data<float>();
-//   at::Tensor at_tensor(c2_tensor);
-//   ASSERT_ANY_THROW(at::sum(at_tensor));
-// }
-
-TEST(Caffe2ToPytorch, PartiallyInitialized) {
-  // These APIs for partially initialized tensors should go away soon, in the
-  // meantime ensure they are caught
-  {
-    // no dtype, no storage
-    caffe2::Tensor c2_tensor(caffe2::CPU);
-    ASSERT_ANY_THROW(at::Tensor at_tensor(c2_tensor));
-  }
-  {
-    // storage, no dtype
-    caffe2::Tensor c2_tensor(caffe2::CPU);
-    c2_tensor.Resize(4,4);
-    ASSERT_ANY_THROW(at::Tensor at_tensor(c2_tensor));
-  }
-  {
-    // dtype, no storage
-    caffe2::Tensor c2_tensor(caffe2::CPU);
-    c2_tensor.Resize(4,4);
-    c2_tensor.mutable_data<float>();
-    c2_tensor.FreeMemory();
-    ASSERT_ANY_THROW(at::Tensor at_tensor(c2_tensor));
-  }
+TEST(TestTensorInterop, Caffe2ToPytorchUnsupportedDevice) {
+  caffe2::Tensor c2_tensor(caffe2::IDEEP);
+  at::Tensor at_tensor(c2_tensor.getIntrusivePtr());
+  ASSERT_ANY_THROW(at::sum(at_tensor));
 }
 
-TEST(PytorchToCaffe2, Op) {
+TEST(TestTensorInterop, PytorchToCaffe2Op) {
   caffe2::Workspace workspace;
   caffe2::NetDef net;
 
@@ -88,13 +63,14 @@ TEST(PytorchToCaffe2, Op) {
   auto at_tensor_b = at::ones({5, 5}, at::dtype(at::kFloat));
   auto at_tensor_c = at::ones({5, 5}, at::dtype(at::kFloat));
 
-  auto* c2_tensor_a = BlobSetTensor(workspace.CreateBlob("a"), caffe2::Tensor(at_tensor_a));
-  auto* c2_tensor_b = BlobSetTensor(workspace.CreateBlob("b"), caffe2::Tensor(at_tensor_b));
+  auto* c2_tensor_a = BlobSetTensor(workspace.CreateBlob("a"), at_tensor_a.getIntrusivePtr());
+  auto* c2_tensor_b = BlobSetTensor(workspace.CreateBlob("b"), at_tensor_b.getIntrusivePtr());
 
   // Test Alias
   {
-    caffe2::Tensor c2_tensor_from_aten(at_tensor_c);
+    caffe2::Tensor c2_tensor_from_aten(at_tensor_c.getIntrusivePtr());
     BlobSetTensor(workspace.CreateBlob("c"), c2_tensor_from_aten.Alias());
+
   }
 
   {
@@ -114,19 +90,19 @@ TEST(PytorchToCaffe2, Op) {
   for (int64_t i = 0; i < 25; i++) {
     ASSERT_EQ(it[i], 3.0);
   }
-  at::Tensor at_result(result);
+  at::Tensor at_result(result.getIntrusivePtr());
   ASSERT_EQ(at::sum(at_result).item<float>(), 75);
 }
 
-TEST(PytorchToCaffe2, SharedStorageRead) {
+TEST(TestTensorInterop, PytorchToCaffe2SharedStorage) {
   caffe2::Workspace workspace;
   caffe2::NetDef net;
 
   auto at_tensor_a = at::ones({5, 5}, at::dtype(at::kFloat));
   auto at_tensor_b = at_tensor_a.view({5, 5});
 
-  auto* c2_tensor_a = BlobSetTensor(workspace.CreateBlob("a"), caffe2::Tensor(at_tensor_a));
-  auto* c2_tensor_b = BlobSetTensor(workspace.CreateBlob("b"), caffe2::Tensor(at_tensor_b));
+  auto* c2_tensor_a = BlobSetTensor(workspace.CreateBlob("a"), at_tensor_a.getIntrusivePtr());
+  auto* c2_tensor_b = BlobSetTensor(workspace.CreateBlob("b"), at_tensor_b.getIntrusivePtr());
 
   {
     auto op = net.add_op();
@@ -143,115 +119,23 @@ TEST(PytorchToCaffe2, SharedStorageRead) {
   for (int64_t i = 0; i < 25; i++) {
     ASSERT_EQ(it[i], 2.0);
   }
-  at::Tensor at_result(result);
+  at::Tensor at_result(result.getIntrusivePtr());
   ASSERT_EQ(at::sum(at_result).item<float>(), 50);
 }
 
-TEST(PytorchToCaffe2, SharedStorageWrite) {
-  auto at_tensor_a = at::ones({5, 5}, at::dtype(at::kFloat));
-  auto at_tensor_b = at_tensor_a.view({25});
-
-  caffe2::Tensor c2_tensor_a(at_tensor_a);
-  caffe2::Tensor c2_tensor_b(at_tensor_b);
-
-  // change is visible everywhere
-  c2_tensor_a.mutable_data<float>()[1] = 123;
-  ASSERT_EQ(c2_tensor_b.mutable_data<float>()[1], 123);
-  ASSERT_EQ(at_tensor_a[0][1].item().to<float>(), 123);
-  ASSERT_EQ(at_tensor_b[1].item().to<float>(), 123);
-}
-
-TEST(PytorchToCaffe2, MutualResizes) {
-  auto at_tensor = at::ones({5, 5}, at::dtype(at::kFloat));
-
-  caffe2::Tensor c2_tensor(at_tensor);
-
-  // change is visible
-  c2_tensor.mutable_data<float>()[0] = 123;
-  ASSERT_EQ(at_tensor[0][0].item().to<float>(), 123);
-
-  // resize PT tensor in smaller direction - storage is preserved
-  at_tensor.resize_({4, 4});
-  c2_tensor.mutable_data<float>()[1] = 234;
-  ASSERT_EQ(at_tensor[0][1].item().to<float>(), 234);
-
-  // resize PT tensor in larger direction - storage is preserved
-  at_tensor.resize_({6, 6});
-  c2_tensor.mutable_data<float>()[2] = 345;
-  ASSERT_EQ(at_tensor[0][2].item().to<float>(), 345);
-  ASSERT_EQ(c2_tensor.sizes()[0], 6);
-  ASSERT_EQ(c2_tensor.sizes()[1], 6);
-
-  // resize Caffe2 tensor - semantics are to NOT preserve the data, but the
-  // TensorImpl is still shared
-  c2_tensor.Resize(7, 7);
-  c2_tensor.mutable_data<float>()[3] = 456;
-  ASSERT_EQ(at_tensor[0][3].item().to<float>(), 456);
-  ASSERT_EQ(at_tensor.sizes()[0], 7);
-  ASSERT_EQ(at_tensor.sizes()[1], 7);
-}
+TEST(TestTensorInterop, PytorchToCaffe2Strided) {
+  caffe2::Workspace workspace;
+  caffe2::NetDef net;
 
-TEST(PytorchToCaffe2, Strided) {
   auto at_tensor = at::ones({5, 5}, at::dtype(at::kFloat)).t();
-  ASSERT_ANY_THROW(caffe2::Tensor c2_tensor(at_tensor));
-  // but calling contiguous is fine
-  caffe2::Tensor c2_tensor(at_tensor.contiguous());
-  for (int64_t i = 0; i < 25; i++) {
-    ASSERT_EQ(c2_tensor.data<float>()[i], 1.0);
-  }
-}
+  auto* c2_tensor = BlobSetTensor(workspace.CreateBlob("blob"), at_tensor.getIntrusivePtr());
 
-TEST(PytorchToCaffe2, InplaceStrided) {
-  auto at_tensor = at::zeros({2, 5}, at::dtype(at::kFloat));
-  caffe2::Tensor c2_tensor(at_tensor);
-  ASSERT_EQ(c2_tensor.sizes()[0], 2);
-  ASSERT_EQ(c2_tensor.sizes()[1], 5);
-
-  c2_tensor.mutable_data<float>()[1] = 234;
-  ASSERT_EQ(at_tensor[0][1].item().to<float>(), 234);
-
-  at_tensor.t_();
-  ASSERT_EQ(c2_tensor.sizes()[0], 5);
-  ASSERT_EQ(c2_tensor.sizes()[1], 2);
-  // This is BROKEN situation, however checking is_contiguous on every data
-  // access is expensive. We rely on user to not do crazy stuff.
-  ASSERT_EQ(at_tensor[1][0].item().to<float>(), 234);
-  ASSERT_EQ(c2_tensor.data<float>()[1], 234);
-}
-
-TEST(PytorchToCaffe2, NonRegularTensor) {
-  at::Tensor at_tensor =
-      at::empty({2, 3}, at::dtype<float>().layout(at::kSparse));
-  ASSERT_TRUE(at_tensor.is_sparse());
-  ASSERT_ANY_THROW(caffe2::Tensor c2_tensor(at_tensor));
-}
-
-// With current build system it's too bothersome to set it up, but the test
-// passes
-// TEST(PytorchToCaffe2, Variable) {
-//   at::Tensor var =
-//       torch::autograd::make_variable(at::empty({2, 3}, at::dtype<float>()));
-//   ASSERT_TRUE(var.is_variable());
-//   ASSERT_ANY_THROW(caffe2::Tensor c2_tensor(var));
-// }
-
-TEST(Caffe2ToPytorch, NonPOD) {
-  caffe2::Tensor c2_tensor = caffe2::empty({1}, at::dtype<std::string>());
-  auto data = c2_tensor.mutable_data<std::string>();
-  *data = "test";
-  ASSERT_ANY_THROW(at::Tensor at_tensor(c2_tensor));
-}
-
-TEST(Caffe2ToPytorch, Nullptr) {
-  caffe2::Tensor c2_tensor;
-  ASSERT_FALSE(c2_tensor.defined());
-  at::Tensor at_tensor(c2_tensor);
-  ASSERT_FALSE(at_tensor.defined());
-}
+  {
+    auto op = net.add_op();
+    op->set_type("Sum");
+    op->add_input("blob");
+    op->add_output("out");
+  }
 
-TEST(PytorchToCaffe2, Nullptr) {
-  at::Tensor at_tensor;
-  ASSERT_FALSE(at_tensor.defined());
-  caffe2::Tensor c2_tensor(at_tensor);
-  ASSERT_FALSE(c2_tensor.defined());
+  ASSERT_ANY_THROW(workspace.RunNetOnce(net));
 }
diff --git a/aten/src/THC/generic/THCTensorMathPointwise.cu b/aten/src/THC/generic/THCTensorMathPointwise.cu
index 82fdd2ac1d..e3ef097515 100644
--- a/aten/src/THC/generic/THCTensorMathPointwise.cu
+++ b/aten/src/THC/generic/THCTensorMathPointwise.cu
@@ -241,7 +241,7 @@ void THCTensor_(cadd)(THCState *state, THCTensor *self_, THCTensor* src1, scalar
 #else
   auto alpha = value;
 #endif
-  at::add_out(out, at::Tensor(retainTensorImpl(src1)), at::Tensor(retainTensorImpl(src2)), alpha);
+  at::add_out(out, retainTensorImpl(src1), retainTensorImpl(src2), alpha);
 }
 
 void THCTensor_(csub)(THCState *state, THCTensor *self_, THCTensor* src1, scalar_t value, THCTensor *src2)
diff --git a/aten/tools/run_tests.sh b/aten/tools/run_tests.sh
index 8ed5a21e2d..c2a0d2f47f 100755
--- a/aten/tools/run_tests.sh
+++ b/aten/tools/run_tests.sh
@@ -35,9 +35,6 @@ fi
 if [[ -x ./cuda_optional_test ]]; then
   ./cuda_optional_test
 fi
-if [[ -x ./cuda_tensor_interop_test ]]; then
-  ./cuda_tensor_interop_test
-fi
 if [ "$VALGRIND" == "ON" ]
 then
   valgrind --suppressions="$VALGRIND_SUP" --error-exitcode=1 ./basic "[cpu]"