Hello,

I found an unexpected behavior of my code in Gem5. The runtime depends of
the starting position of the code.

I'm running on ARM cpu for the detailed processor model compiled with opt
flavor.

I made an experiment to exemplify this situation. It consists of running a
code for different alignments.

The results are:

case         end              init         diff
00    18233000    13203000    5030000
01    18233000    13203000    5030000
02    18233000    13203000    5030000
03    18233000    13203000    5030000
04    18290000    13208000    5082000
05    18284500    13208000    5076500
06    18284500    13208500    5076000
07    18285500    13208500    5077000
08    23285000    13208500    10076500
09    23285500    13208500    10077000
10    23285500    13208500    10077000
11    23299000    13208500    10090500
12    23285500    13208500    10077000
13    23285500    13208500    10077000
14    18286500    13209000    5077500
15    18304000    13274000    5030000

The first column is the value for the alignment. The two nexts are the gem5
ticks for end and start. The last one is the difference.

As you can see, there is a set of alignments that costs twice. It seems
something is not working properly.

In addition, I count the ifetch events for miss and hit, and I found the
following results,

case    miss    hit
00    357    1440
01    357    1440
02    357    1440
03    357    1440
04    357    1441
05    357    1436
06    357    1437
07    357    1437
08    357    11436
09    357    11437
10    357    11437
11    359    11431
12    357    11435
13    357    11435
14    357    1436
15    358    1440

The amount of misses is the same in all the cases, but the amount of hits
is 10 times more for some of them.

I think, that the problem is on the pipeline of out-of-order. I checked the
same experiment using the minor cpu model, and it works ok.

I'm a file containing all the necessary files to reproduce the experiment.
I tested it on the last repo version for stable and developer.

Thanks in advance,
David
_______________________________________________
gem5-dev mailing list
[email protected]
http://m5sim.org/mailman/listinfo/gem5-dev

Reply via email to