Function AreCastCompatible

tensorflow/compiler/mlir/tensorflow/ir/tf_ops.cc:115–152

Returns true if the given pair of TensorFlow types can be cast to one another. In other words, a single run-time value is legal for both the types. For example, tensor<*xf32> and tensor<3xf32> are cast compatible.

// another. In other words, a single run-time value is legal for both the types.
// For example, tensor<*xf32> and tensor<3xf32> are cast compatible.
static bool AreCastCompatible(Type a, Type b) {
  if (TensorCastOp::areCastCompatible(a, b)) return true;

  // Resource types may optionally contain subtypes information that does not
  // match. Check subtypes compatibility when possible, otherwise treat them as
  // compatible.
  auto a_or_element_type = getElementTypeOrSelf(a);
  auto b_or_element_type = getElementTypeOrSelf(b);

  auto a_kind = a_or_element_type.getKind();
  auto b_kind = b_or_element_type.getKind();

  if (a_kind == TensorFlowTypes::RESOURCE &&
      b_kind == TensorFlowTypes::RESOURCE) {
    auto a_resource_type = a_or_element_type.dyn_cast<ResourceType>();
    auto b_resource_type = b_or_element_type.dyn_cast<ResourceType>();
    bool a_has_subtype = !a_resource_type.getSubtypes().empty();
    bool b_has_subtype = !b_resource_type.getSubtypes().empty();

    if (!a_has_subtype || !b_has_subtype) return true;

    assert(a_resource_type.getSubtypes().size() <= 1 &&
           "Resource type must have at most one subtype");
    assert(b_resource_type.getSubtypes().size() <= 1 &&
           "Resource type must have at most one subtype");

    return TensorCastOp::areCastCompatible(
        a_resource_type.getSubtypes().front(),
        b_resource_type.getSubtypes().front());
  }

  // Variant types may optionally contain subtypes information that need not
  // match.  It is also not possible to compare subtypes for compatibility as
  // their interpretation depends on the ops operating on them. So, accept all
  // pairs of variant types.
  return a_kind == TensorFlowTypes::VARIANT &&
         b_kind == TensorFlowTypes::VARIANT;
}

static bool IsUnknownDimOrRank(int64_t dim_or_rank) {
  return dim_or_rank == -1;