Which is the backward output of a bidirectional RNN,and how can i get the backward result

Continuing the discussion from How can I know which part of h_n of bidirectional RNN is for backward process?:

I really want to make sure which is right
[
layer0_forward
layer0_backward
layer1_forward
layer1_backward
layer2_forward
layer2_backward

]
or
[
layer0_forward
layer1_forward
layer2_forward
layer0_backward
layer1_backward
layer2_backward

]

if first is right,how can i get it?

Is there anyone knows how the composition of the rnn output?
such as,
rnn = nn.lstm( 100, 100, layers = 3, bidirectional=True )
output = rnn(input)
what is the composition of the output?