Configure and flint-mparam for meteorlake

[vneiger]

This allows Intel's meteorlake CPUs to be identified during configure. This adds corresponding parameters in meteorlake/flint-mparam.h.

[vneiger]

This made me realise that recent Intel processors (RaptorLake, AlderLake, ... probably all those more recent than IceLake) will not benefit from the enhancement in #2233 , since they still rely on SkyLake's flint-mparam.h. Should we create a flint-mparam for each of them? Or, more simply (but maybe less cleanly), should we make them all rely on IceLake's parameters, since they coincide for all we know at the moment?

[albinahlback]

These would be nice to figure out, but I suppose they are conservative at best. Regardless, we need a proper tuning suite as it is quite tedious to do these by hand.

[vneiger]

Yes, I was not sure how to find them easily (i.e. without looking through the code to find out what they mean, then create some profile/tune file, etc.). Most others were output by a tuning file (FFT_TAB, MULMOD_TAB, etc.).

[vneiger]

The latest commits added a flint-mparam file for Tiger Lake and Raptor Lake. These commits also make other Intel processors from architectures more recent than SkyLake fall back to IceLake's flint-mparam instead of SkyLake. This way these processors will benefit from the change in #2233 , and also have a more relevant threshold for modular multiplications with precomputations (mulmod_shoup). Both aspects boil down to the efficiency of udiv_qrnnd, and since I didn't have access to these processors I checked on uops.info that they should behave similarly to IceLake.

[vneiger]

@albinahlback a quick re-check of the recent commits could be useful before merging.

[albinahlback]

I think it looks good!

[albinahlback]

Sorry, I was about to squash and merge the commits. I only discussed this with Fredrik, so it is my fault for not bringing this forward to everyone:

Every PR should now be squashed when merging.

This allows us to:

• Merge all commits into one single commit.
• The Git log now contains the number of the pull request in its title.
• Avoid duplicate commit messages by amending the Git log. In this case, two commits have the title fallback to icelake instead of skylake.
• Amend the Git log to make it more descriptive. To me, add raptorlake and tigerlake is not that descriptive. Of course, I know what it means in this context, but looking back at this in say 1 year, I would probably have to look at the difference to tell what it really meant. Probably something like Add flint-mparam.h for raptorlake and tigerlake would be more suitable.

Now, I took examples from here for convenience, but please do not feel targeted.

@fredrik-johansson @vneiger @edgarcosta

[fredrik-johansson]

What Albin said, to make the history easier to follow, though I would relax this:

Every PR should now be squashed when merging.

Not every PR necessarily; some PRs do several related things, and it's fine to keep the individual commits in some cases if they make sense independently.

[vneiger]

Sorry, I was about to squash and merge the commits

Oops, ok. I had seen a comment of this kind in some recent PR or issue, but maybe did not pay close enough attention to it, sorry.

Thanks for explaining the recommended procedure in detail!