Hey Ali, I sent the patch through "hg email" to the dev mailing list.
I've got another suggestion for the tournament predictor: maybe we should drop some parameters we can initialize the predictor with. Since the same threshold is used for all 3 predictors (global-, local- and choice-predictor) some parameters are dependent. An example: choiceCtrBits and globalCtrBits should both be equal to localCtrBits (in cpu/pred/tournament.{cc,hh}). These dependencies can be omitted when we use 3 different thresholds. However there are still some other dependencies that can't.
It might be better for someone who wants to set parameters for the branch predictor, he doesn't need to take these dependencies into account.
Max On 05/03/2010 07:44 PM, Ali Saidi wrote:
Hi Max, I think that is a bug. Would you mind working up a diff, comment, etc. I'll volunteer to update all of our local regressions if you do. Thanks, Ali On May 3, 2010, at 4:28 AM, Maximilien Breughe wrote:Hi all, I was digging a little bit in the implementations of your branch predictors and I think I bumped into a small bug (which might affect performance of the target machine a lot): Around line 108 in cpu/pred/tournament.cc: threshold = (1<< (localCtrBits - 1)) - 1; threshold = threshold / 2; Shouldn't this be threshold = (1<< (localCtrBits - 1)) - 1; ? When you take localCtrBits=2 for example, the threshold should be 1, not 0. Regards, Max _______________________________________________ m5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/m5-users_______________________________________________ m5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
_______________________________________________ m5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
