About the runtime of archived TritonRoute and TritonRoute in OpenROAD
Tool: Detailed Routing
Subcategory: Performance degradation
Conversation
gilgamsh
I have tried both of them and running the same ispd19 benchmark. I have noticed that the runtime of TritonRoute in OpenROAD is quite larger than that in archived TritonRoute. The specific data is In ispd19_testcase1 , with single thread TritonRoute in OpenROAD : Elapsed Time: 643.091s archived TritonRoute. :Elapsed Time: 236.032s
I have noticed that, in TritonRoute in OpenROAD ,the pin analysis time and design rule checking time increase dramatically. pin analysis time in OpenROAD, about 300s, pin analysis time in archived TritonRoute, about 10s. Is there a normal situation? Maybe there is something wrong about my scripts.
I use the openroad scripts like below
read_lef ../benchmarks//ispd2019/ispd19_test1/ispd19_test1.input.lef
read_def ../benchmarks//ispd2019/ispd19_test1/ispd19_test1.input.def
read_guides ../benchmarks//ispd2019/ispd19_test1/ispd19_test1.guide
detailed_route
write_def ispd19_test1.solution.def
exit
gilgamsh
another question, If i want to learn the code of TritonRoute-WXL: The Open Source Router with Integrated DRC Engine. Can i use the archived TritonRoute, thinking it implemented the functions mentioned in this paper(not including the gr)?
maliberty
Have you enabled a similar number of threads in both runs? In OR use set_thread_count <n>
The archived version should match the paper but any further improvements will not be present.
maliberty
That should be fine. Are you using a similar number of threads in standalone TR?
antonblanchard
Are you running a recent version of OpenROAD with the following fix: https://github.com/The-OpenROAD-Project/OpenROAD/commit/0cf19963d358347bcf06a2cd803ea883f8c46f27
That issue isn't in the original TritonRoute source, and can cause significantly more runtime in the pin access code.
antonblanchard
@vijayank88 did you mean to close this? I just checked, and we are spending significantly more time in genInstRowPattern vs TritonRoute.
antonblanchard
Most of the problem appears to be 5fada4c44b2d4aba09185eb1217f928645c32c4b. Start up time for ispd19_test1 goes from 17 seconds to 280 seconds. I haven't got far enough to understand if all that overhead is required. FYI @maliberty @osamahammad21
maliberty
Its not obvious to me how that change triggers the issue but its @osamahammad21 change so he might have an idea.
antonblanchard
I'm not sure why, but the problem is caused by filtering out other instances when generating our preferred access points:
- hasVio = !genPatterns_gc({prevInst, currInst}, objs, Edge);
+ // hasVio = !genPatterns_gc({prevInst, currInst}, objs, Edge);
+ hasVio = !genPatterns_gc({}, objs, Edge);
I would have expected this to take less time, unless the act of filtering out other instances is taking lots of time.
oharboe
@gilgamsh It would be interesting to see a profile of both cases to see where the difference in time is. That should tell us something of the nature of the difference to corroborate the suspicions on what differences in code that can cause this.
maliberty
Note that a fair number of enhancements and bug fixes have gone into the OR version. It is possible you are not observing equivalent results.