MPI_Op_create(user_function, commute, op) MPI_Op_free(op)
user_function(invec, inoutvec, len, datatype)
用户函数应执行
inoutvec[i] = invec[i] op inoutvec[i]; for i from 0 to len-1.
user_function可以是非交换的(例如,矩阵乘)