Daily GPU tests fail to output the random seed used about gem5 HOT 6 CLOSED

Why do we have non-determinism anywhere in gem5? To me this isn't a problem with the tests, there shouldn't be anything truely random going on. The "time-based seed" should be replaced with something static.

from gem5.

Yeah, I was thinking of just altering the files in the tests directory to use --random-seed= so that we don't have to make more changes to the workflow directory. If that's okay, I can just assign this to myself and get that done.

from gem5.

I had asked @DanKouch to do the same @mkjost0 , but if you want to get to it today, please go ahead.

@abmerop (or really, Brad) will have to comment more on @BobbyRBruce 's original question though, as I was not involved with that work.

from gem5.

Changing the command line parameter is the best thing to do in the short term (I think).

Longer term, the problem is this line:

We should be using mt_random here instead of srand. We should always use gem5's random number generator so that we have determinism by default.

Then, when we want to have multiple different seeds used, we can use (from python) m5._m5.core.seedRandom(random.randint(0,100000)) (or something like this). We should expose this seedRandom function better, though.

Anyway, the "right" solution is a bit deeper than is worth it at this point. For now, let's just set the command line parameter.

from gem5.

Changing the command line parameter is the best thing to do in the short term (I think).

Longer term, the problem is this line:

gem5/src/cpu/testers/gpu_ruby_test/protocol_tester.cc

Line 148 in af72b9b

srand(time(NULL));

We should be using mt_random here instead of srand. We should always use gem5's random number generator so that we have determinism by default.

Then, when we want to have multiple different seeds used, we can use (from python) m5._m5.core.seedRandom(random.randint(0,100000)) (or something like this). We should expose this seedRandom function better, though.

Anyway, the "right" solution is a bit deeper than is worth it at this point. For now, let's just set the command line parameter.

I had a go at this in #140. I'm not an expert on this part of the codebase, but I think my implementation of random_mt works.

from gem5.

@mkjost0, can you check the status of this since #140 has been merged?

from gem5.