By: Anon (no.delete@this.spam.com), April 2, 2021 10:51 am
Room: Moderated Discussions
Carson (carson.delete@this.example.edu) on April 1, 2021 11:13 pm wrote:
> If the probability of success (in a given prediction context) is less than
> this, the predictor should skip HTM and go straight to the fallback path.
>
> It's not at all clear that this cost ratio can be approximated by a fixed
> constant, even within the factor-of-2 accuracy of a cheap predictor.
>
> Which in turn means that the SW tell the HW when the fallback path completes so that C(fallback)
> can be measured. That's a programmer-visible requirement that Intel's RTM doesnt include.
Just to know, why don't put all the fallback in hardware? Or, if I understand correctly how people call it, garantee forward progress?
This half hardware half software solution seens to always falls in the problem of both halves lacking information about the other half.
> If the probability of success (in a given prediction context) is less than
> this, the predictor should skip HTM and go straight to the fallback path.
>
> It's not at all clear that this cost ratio can be approximated by a fixed
> constant, even within the factor-of-2 accuracy of a cheap predictor.
>
> Which in turn means that the SW tell the HW when the fallback path completes so that C(fallback)
> can be measured. That's a programmer-visible requirement that Intel's RTM doesnt include.
Just to know, why don't put all the fallback in hardware? Or, if I understand correctly how people call it, garantee forward progress?
This half hardware half software solution seens to always falls in the problem of both halves lacking information about the other half.