why compute ops like conv serially, but not parallelly

[ddt2014]

For example, in src/operator/nn/convolution-inl.h 307, mxnet compute the conv operator serially for each batch, which leading to the forward time(batch size > 1) much longer than batch size == 1.

I wanna know the reason of this

[frankfliu]

@mxnet-label-bot add [operator, question]