By: --- (---.delete@this.redheron.com), August 28, 2022 2:16 pm
Room: Moderated Discussions
Kara (karaardalan.delete@this.gmail.com) on August 27, 2022 1:00 pm wrote:
>
> Someone said if the branch pattern is easily predicted,
> or has a clear patter, then let the hardware do it :)
>
> Bruh :) how lazy would you have to be.
>
> If it's easy, you do it. Tell the compiler where to look, use your
> avx instructions my friend, use the big juicy regs on there.
>
> And if the branching have chaotic directions, then let's do everyone a favor
> and not try to predict it by a perceptron algorithm of things, shall we?
>
Hashed perceptron based prediction, as long as it uses multiple components of different lengths, is remarkably good; not quite as good as TAGE, but close.
Current state of the art is discussed here:
https://people.engr.ncsu.edu/hzhou/CNN_DBN_zhou_2017.pdf
and while likely impractical today, NN-based prediction may yet return to its former glory.
Of course if your branch is truly uncorrelated with the past, the best you can do is a statistical predictor; if it's truly uncorrelated with the past AND 50% taken then past-based prediction is hopeless. BUT note those *precise* words...
Branches that are uncorrelated with past branches may be predictable via alternative means that do not rely on statistics from the past, and that is one of the newer frontiers of branch research, discovering those cases and exploiting them: https://users.elis.ugent.be/~leeckhou/papers/hpca2020.pdf
>
> Someone said if the branch pattern is easily predicted,
> or has a clear patter, then let the hardware do it :)
>
> Bruh :) how lazy would you have to be.
>
> If it's easy, you do it. Tell the compiler where to look, use your
> avx instructions my friend, use the big juicy regs on there.
>
> And if the branching have chaotic directions, then let's do everyone a favor
> and not try to predict it by a perceptron algorithm of things, shall we?
>
Hashed perceptron based prediction, as long as it uses multiple components of different lengths, is remarkably good; not quite as good as TAGE, but close.
Current state of the art is discussed here:
https://people.engr.ncsu.edu/hzhou/CNN_DBN_zhou_2017.pdf
and while likely impractical today, NN-based prediction may yet return to its former glory.
Of course if your branch is truly uncorrelated with the past, the best you can do is a statistical predictor; if it's truly uncorrelated with the past AND 50% taken then past-based prediction is hopeless. BUT note those *precise* words...
Branches that are uncorrelated with past branches may be predictable via alternative means that do not rely on statistics from the past, and that is one of the newer frontiers of branch research, discovering those cases and exploiting them: https://users.elis.ugent.be/~leeckhou/papers/hpca2020.pdf