Determine whether you should split the design into multiple kernels connected by pipes.
Aggregate data on pipes only when all the data is used at the same point in the kernel.
Do not use non-blocking pipes if you are using a looping structure waiting for the data, that is, avoid the following coding pattern for non-blocking pipe accessors:

bool success = false;
while (!success) {
  my_pipe::write(rd_src_buf[i], success); // can be a non-blocking read too
}