Complex loop exit condition

Your inner loop has a variable exit condition; hence, the outer loop is not pipelineable. If ei has a maximum limit, you can replace ei in the header of the inner loop with that maximum limit, and add a branch inside of the inner loop to discard cases where j >= ei. That will allow pipelining of both loops, at cost of redundant computation which could be significant depending on how far ei is from that maximum. If there is no maximum or it is too large, then an NDRange kernel would be more appropriate since the thread scheduler can minimize pipeline stalls bubbles at run-time.