Note [Undefined-dim versus 0-dim]

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

Author: ezyang

templates/TensorDerived.cpp

@@ -29,6 +29,7 @@ int64_t ${Tensor}::dim() {
  int64_t d = ${THTensor}_nDimension(${state,}tensor);
  if(d != 0)
  return d;
+ // See Note [Undefined-dim versus 0-dim]
  return kUndefinedDimensions;
 }
 

templates/Type.h

@@ -46,6 +46,16 @@ constexpr Backend kCUDA = Backend::CUDA;
 constexpr Backend kSparseCPU = Backend::SparseCPU;
 constexpr Backend kSparseCUDA = Backend::SparseCUDA;
 
+// Note [Undefined-dim versus 0-dim]
+// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+// Unlike Torch, ATen treats zero-dimension tensors as having ONE
+// element (that is to say, a zero-dimensional tensor is a scalar!)
+// This is in contrast to Torch, where a zero-dimension tensor has
+// zero elements.
+//
+// Because we are backed by Torch tensors, we need to be able to
+// represent this state (of numel==0). kUndefinedDimensions represents this
+// situation.
 constexpr int64_t kUndefinedDimensions = std::numeric_limits<int64_t>::min();
 
 static inline const char * toString(Backend b) {