// Returns number of elements across all dimensions.
  virtual Status NumElements(int64_t* num_elements) const = 0;
  // Returns size of specified dimension
  //
  // -1 indicates an unknown axis length; this is unreachable for most standard
  // ImmediateExecutionTensorHandles, but comes up for example when computing
  // the shape of a parallel tensor with component shapes differing across
  // devices.

  TFE_TensorHandle* axis = TestAxisTensorHandle(ctx);
  TFE_Op* minOp = MinOp(ctx, input, axis);
  TFE_TensorHandle* retvals[1] = {nullptr};
  int num_retvals = 1;
  TFE_Execute(minOp, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteOp(minOp);
  TFE_DeleteTensorHandle(input);
  TFE_DeleteTensorHandle(axis);
  ASSERT_EQ(1, num_retvals);

  // Make a placeholder operation.
  TF_Operation* feed = Placeholder(graph, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Make a constant operation with the scalar "0", for axis.
  TF_Operation* one = ScalarConst(0, graph, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Create a session for this graph.
  CSession csession(graph, s, use_XLA);

      } else {
        // Generalize differing axis lengths to "variable"/"unknown".
        for (int axis_index = 0; axis_index < combined_shape.dims();
             ++axis_index) {
          int64_t axis_length = combined_shape.dim_size(axis_index);
          if (axis_length != component_shape.dim_size(axis_index)) {
            axis_length = -1;
          }
          TF_RETURN_IF_ERROR(

// Return a 1-D INT32 tensor containing a single value 1.
TFE_TensorHandle* TestAxisTensorHandle(TFE_Context* ctx);

// Return an op taking minimum of `input` long `axis` dimension.
TFE_Op* MinOp(TFE_Context* ctx, TFE_TensorHandle* input,
              TFE_TensorHandle* axis);

// If there is a device of type `device_type`, returns true
// and sets 'device_name' accordingly.
// `device_type` must be either "GPU" or "TPU".

        // Non-parallel tensors from _EagerConst/tf.constant are implicitly
        // broadcast, i.e. set as the input to each parallel operation. This
        // allows code like "tf.constant(1.)" or "tf.reduce_sum(..., axis=1)"
        // (where the value starts on the host), without allowing other implicit
        // copies/broadcasts. Other implicit copies may be supported eventually,

  // retrieving shapes of tensors they wrap until the custom device tensor's
  // shape is explicitly requested where possible.
  int (*num_dims)(void* data, TF_Status* status);

  // Computes the axis length at `dim_index`.
  int64_t (*dim)(void* data, int dim_index, TF_Status* status);

  void (*deallocator)(void* data);

  // Summarizes the value of this tensor. The caller takes ownership of the

TFE_Op* MinOp(TFE_Context* ctx, TFE_TensorHandle* input,
              TFE_TensorHandle* axis) {
  TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "Min", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, input, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, axis, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

`tf.reduce_join`: `reduction_indices` becomes `axis` * `tf.reduce_logsumexp`:
`reduction_indices` becomes `axis` * `tf.reduce_max`: `reduction_indices`
becomes `axis` * `tf.reduce_mean`: `reduction_indices` becomes `axis` *
`tf.reduce_min`: `reduction_indices` becomes `axis` * `tf.reduce_prod`:
`reduction_indices` becomes `axis` * `tf.reduce_sum`: `reduction_indices`