Hi, when I use bitserial conv2d or dense operators for ARM CPUs I always get alignment warnings. For the convolution operator:
Warning: Trying to bind buffer to another one with lower alignment requirement required_alignment=128, provided_alignment=64
and for the dense operator:
Warning: Trying to bind buffer to another one with lower alignment requirement required_alignment=128, provided_alignment=8
I suppose it’s harmful to performance not ignore it or isn’t it that important? Can I do anything about it on the relay level? Is there an parameter I need to set when creating the weights or the input values? Till yet I do it like this:
input_array = relay.var( ‘data’, shape=(1,32,32,32), dtype=‘int16’)
weight = relay.var(‘conv2d_weight’, relay.TensorType( (3,3,32,32), ‘uint32’))
@jwfromm, maybe I have to ask you what is the right way to deal with this. Thank you!