Example for conv2d: ComputationCost = ((k*k + k*k-1)*c + c-1 + bias?1:0) * out_channel * out_width * out_height * batch_size = (2*k*k*c - 1 + bias?1:

generate a temporary object to allow `const auto& x = optval().value()` where `optval()` is a function call which returns a temporary Optional

Drop down the maximum layer with the minimum layer form consumer

Drop down the maximum layer with the minimum layer form consumer

Compute the cost between this node and adjacent nodes with a lower order

Judge whether an NdSbp could be applied on a tensor with given logical shape True means this NdSbp is not valid.

Theoretically computation cost of matrix multiplication is the products of the number of matrix and first dimension of matrix a, second dimension of m

Get the minimum element in Cost

use this func to union the ptr type in this conversion phase.

NOTE: Get random seed in current rank, and ensure that it will have same seed between broadcast sbp and it will be different between split sbp. It wi

[start, end)

Construct a boxing collector with given maximum number of axis