Sparse tensor core for convolution requires the input channel dimension is 2:4 structed sparse.
So we have to mask the input channel dimension for using sparse tensor core