Re: [Gems-users] Help with Opal Tester

[Ankit Jalote <jalote@xxxxxxxxxx>]

Hi,

Regarding the first option : I think there is a problem. The trace may 
not match the opal simulation
because of the OS activity going on (hadling interrupts) as the sequence 
of instructions being fetched is different. Do you think
this problem can be avoided?
What is was thinking was that if "tester" runs the trace exactly as the 
opal simulations except ofcourse it is reading from the trace then that
should effectively be a perfect branch predictor. But I am not sure if 
tester was written for this purpose. Can you please let me know what the 
tester was designed for?

Thanks,
Ankit

Luke Yen wrote:

> Ankit,
>
>  I have looked into this problem before (trying to do the same thing),
> and I think there are a couple options.
>
> 1) As I have mentioned before there is no perfect BP flag in Opal, and I
> recommended using the Opal trace utility to generate branch traces and
> then feeding it back into Opal.  I am not sure if that is currently
> an option for Opal, or if it just works for the tester.  I can imagine
> someone modifying Opal to read from the branch trace in real time, while
> simulating a workload.
>
> 2) Carl Mauer (author of Opal) was experimenting with some runahead type
> ideas, because Opal has some initial code to construct a control flow
> graph.  You can find this code in pseq.C.  Search for flow_inst_t, which
> are instructions that are linked together to form a CFG.  However one of
> the problems I saw is that in order to construct the CFG one must actually
> step Simics, which will conflict with simulating the OoO timing of the
> processor.  I haven't figured out a solution to this problem, and maybe
> you can figure out how to get this working and post a solution.
>
> 3) I am not sure if MAI has some perfect branch prediction options but
> maybe one can write a simple MAI module that can hook up to Opal whenever
> it sees a branch instruction, and then choose the correct path.
>
>  Regards,
>   Luke
>
> On Fri, 8 Jul 2005, Ankit Jalote wrote:
>
>  
>
> > Hi,
> > Thanks for the reply. I will look at the code to figure this out.
> > I have one additional question though. I am trying to simulate perfect
> > branch prediction in gems. Can you please give me some pointers to do this?
> > One way I can think of is by creating the trace and then running the
> > created trace through tester. But is there a better way to implement
> > perfect prediction?
> > Thanks a lot,
> > Ankit
> >
> >    
> _______________________________________________
> Gems-users mailing list
> Gems-users@xxxxxxxxxxx
> https://lists.cs.wisc.edu/mailman/listinfo/gems-users
>