Revert D25544731: [PyTorch] Avoid extra Tensor refcounting in _cat_ou…

…t_cpu

Test Plan: revert-hammer

Differential Revision:
D25544731 (1a05104)

Original commit changeset: 7b9656d0371a

fbshipit-source-id: 0f7ea74eca282cadf269bbd284d59650a431ed65

aten/src/ATen/native/TensorShape.cpp

@@ -101,10 +101,6 @@ static inline void check_cat_shape_except_dim(const Tensor & first, const Tensor
   }
 }
 
-static bool should_skip(const Tensor& t) {
-  return t.numel() == 0 && t.dim() == 1;
-}
-
 Tensor & _cat_out_cpu(Tensor& result, TensorList tensors, int64_t dim) {
   // previously, size [0] tensors were the only possible empty tensors; thus, it wasn't possible
   // to cat empty tensors unless all the other tensors were 1-dimensional, so we allowed these tensors
@@ -113,6 +109,7 @@ Tensor & _cat_out_cpu(Tensor& result, TensorList tensors, int64_t dim) {
   // FIXME: warn if this is the case
   bool allSkipped = true;
   bool allContiguous = true;
+  Tensor notSkippedTensor;
 
   // Inputs cannot alias the output tensor
   for (int64_t i = 0; i < tensors.size(); i++) {
@@ -124,21 +121,19 @@ Tensor & _cat_out_cpu(Tensor& result, TensorList tensors, int64_t dim) {
   }
   at::assert_no_internal_overlap(result);
 
-  const Tensor* pnotSkippedTensor = [](TensorList tensors) -> const Tensor* {
-    for (auto const &tensor : tensors) {
-      if (should_skip(tensor)) {
-        continue;
-      }
-      // we've found a non-empty tensor
-      return &tensor;
+  auto should_skip = [](const Tensor& t) { return t.numel() == 0 && t.dim() == 1; };
+  for (auto const &tensor : tensors) {
+    if (should_skip(tensor)) {
+      continue;
     }
-    return nullptr;
-  }(tensors);
-
-  if (!pnotSkippedTensor) {
+    // we've found a non-empty tensor
+    allSkipped = false;
+    notSkippedTensor = tensor;
+    break;
+  }
+  if (allSkipped) {
     return result;
   }
-  const Tensor& notSkippedTensor = *pnotSkippedTensor;
 
   TORCH_CHECK(tensors.size() > 0, "expected a non-empty list of Tensors");
   TORCH_CHECK(dim <= notSkippedTensor.dim(), "dimension ", dim, "out of range");
@@ -196,7 +191,7 @@ Tensor & _cat_out_cpu(Tensor& result, TensorList tensors, int64_t dim) {
   if (reuse_iterator &&
       result.is_contiguous(first_tensor_mem_format) &&
       no_type_promotion) {
-    const auto& source_slice = notSkippedTensor;
+    auto source_slice = notSkippedTensor;
     auto slice_dim_size = source_slice.size(dim);
     auto result_slice = result.narrow(dim, 0, slice_dim_size);
     auto result_slice_data = result_slice.data_ptr();