By: David Kanter (dkanter.delete@this.realworldtech.com), November 1, 2008 11:53 pm
Room: Moderated Discussions
Howard Chu (hyc@symas.com) on 10/31/08 wrote:
---------------------------
>Given the flakiness of the tools, it would have been worthwhile to hack the codepath
>selection in each of the programs-under-test, to force identical codepaths on both
>platforms.
I'm not entirely sure how I could do that without faking CPUID, which I'm sure is possible, but I don't know how to do it, and I suspect it might consume a lot of time.
> Then you could assess just how badly the tools were lying. >As it is,
>you have a variable propagating systematically through all >your results and no way to account for it.
Yes and no. I trust the VTune results - I'm sure they are numerically accurate. The user (me) might have screwed something up (i.e. not picking the right event or misinterpreting the meaning). But there was almost never run-to-run variation. Code analyst I'm a lot more nervous about because it did not appear to be numerically stable.
I think that the next time around I may just stick with intel chips, so that I don't have as much uncertainty in the results.
David
---------------------------
>Given the flakiness of the tools, it would have been worthwhile to hack the codepath
>selection in each of the programs-under-test, to force identical codepaths on both
>platforms.
I'm not entirely sure how I could do that without faking CPUID, which I'm sure is possible, but I don't know how to do it, and I suspect it might consume a lot of time.
> Then you could assess just how badly the tools were lying. >As it is,
>you have a variable propagating systematically through all >your results and no way to account for it.
Yes and no. I trust the VTune results - I'm sure they are numerically accurate. The user (me) might have screwed something up (i.e. not picking the right event or misinterpreting the meaning). But there was almost never run-to-run variation. Code analyst I'm a lot more nervous about because it did not appear to be numerically stable.
I think that the next time around I may just stick with intel chips, so that I don't have as much uncertainty in the results.
David