int i, src, dst, rank, num_procs;
int X, Y, send_offset, recv_offset;
int my_row_base, my_col_base, src_row_base, block_size, num_reqs;
int i, src, dst, rank, num_procs;
int X, Y, send_offset, recv_offset;
int my_row_base, my_col_base, src_row_base, block_size, num_reqs;