AMD Developers Looking At GNU C Library Platform Optimizations For Zen

[carewolf]

is this because Zen 1 can only do AVX2 back packing together two 128bit operations to achieve a 256bit AVX2 calc?

i.e. flipping this on by default for AMD architectures willgreatly benefit Zen2, but might actually slow down Zen1.

No it won't. It just won't speed up as much. Note this isn't just AVX2, it also include SSSE3 and SSE4.1 both of which gives much bigger benefits to automatically vectorized code than going from 128bit to 256bits per instruction does.

[ddriver]

Please give concrete examples of this happening in Linux or glibc.

I wasn't referring to neither, dunno where you got that idea. What I was referring to is that they have a long trail of mischief on every fathomable vector, they have abused software, hardware, law, finances, economy ... you name it. That's just intel being intel, and that sets the bar pretty high... or low, depending on how you look at it. Either way, given their history, such an accidental issue you describe will hardly generate outrage, it would be actually quite understandable even for someone like me, who doesn't see intel in even remotely favorable light.

[Space Heater]

I wasn't referring to neither, dunno where you got that idea.

The fact that this whole thread is about CPU optimizations in glibc, and you quoted a post of mine that was specifically talking about Intel contributing to glibc and the kernel.

What I was referring to is that they have a long trail of mischief on every fathomable vector, they have abused software, hardware, law, finances, economy ... you name it. That's just intel being intel, and that sets the bar pretty high... or low, depending on how you look at it.

Except that open source projects have maintainers and they have the final say on whether or not a a patch set is committed to the project. It's one thing for Intel to hurt competitors on software projects wholly owned by Intel, but to claim it's par for the course for Intel to do this in open source projects is frankly a huge insult to the maintainers. You are implicitly suggesting that either the maintainers are incompetent, corrupt, or are easily tricked into accepting whatever patches Intel gives them.

Either way, given their history, such an accidental issue you describe will hardly generate outrage, it would be actually quite understandable even for someone like me, who doesn't see intel in even remotely favorable light.

I seriously doubt that people who already dislike Intel would for some reason not say a word if Intel caused a serious performance regression on AMD systems, they would be justified in loudly complaining too.

[Teggs]

There were regressions from just 'turning it on'. I won't link directly to joebonrichie 's Solus thread, since he didn't, but there were a few issues on Zen 1 and Zen+ that he worked out in particular.

I am in complete agreement that this should and could have been done long since, upstream, but just enabling it with no further testing would have hurt a few workloads. (And maybe that would have gotten attention at the time, but it would have been unprofessional.)

[ms178]

There were regressions from just 'turning it on'. I won't link directly to joebonrichie 's Solus thread, since he didn't, but there were a few issues on Zen 1 and Zen+ that he worked out in particular.

I am in complete agreement that this should and could have been done long since, upstream, but just enabling it with no further testing would have hurt a few workloads. (And maybe that would have gotten attention at the time, but it would have been unprofessional.)

Agreed, but such testing could have been done or required before merging the ifunc functionality in the first place as this feature is not architecture-specific in nature and its purpose was to help to spread the usage of these x86 ISA extensions.

[geearf]

Yeah, that is why it is best to pirate every single Nvidia sponsored game. I find it fair and just, Nvidia has already paid for the product, why should i?

That's actually not a horrible idea.
I've always hated getting advertisement, for Nvidia and others of course, when I start a game I paid full price for, maybe for now I will refund those right away.
Of course credits for the technology used by the game when I play it is fine, and different from advertisement of something I won't actually use.

[mos87]

And that's how free software should work.

[TemplarGR]

That's actually not a horrible idea.
I've always hated getting advertisement, for Nvidia and others of course, when I start a game I paid full price for, maybe for now I will refund those right away.
Of course credits for the technology used by the game when I play it is fine, and different from advertisement of something I won't actually use.

No, in that case it is not advertisement, if it was advertisement we could let it slide. Nvidia pays to make the games perform worse on AMD by providing stuff like "gameworks" which choke the pipeline by abusing the tesselator for example. I am sorry but expecting AMD owners to pay developers who agree to such terms is delusional. This thing in an ideal world would have been illegal to do since it is anti-competitive but we don't live in an ideal world so Nvidia and those dev partners of theirs can get away with it.

Well guess what, it appears that this world is indeed not perfect and they can't catch me pirating their shit either. They are abusing their sponsorship deals to win in benchmarks because they can, and i abuse the torrent sites because i can. LOL.

[geearf]

No, in that case it is not advertisement, if it was advertisement we could let it slide. Nvidia pays to make the games perform worse on AMD by providing stuff like "gameworks" which choke the pipeline by abusing the tesselator for example. I am sorry but expecting AMD owners to pay developers who agree to such terms is delusional. This thing in an ideal world would have been illegal to do since it is anti-competitive but we don't live in an ideal world so Nvidia and those dev partners of theirs can get away with it.

Well guess what, it appears that this world is indeed not perfect and they can't catch me pirating their shit either. They are abusing their sponsorship deals to win in benchmarks because they can, and i abuse the torrent sites because i can. LOL.

Oh I see, thank you!

[finalzone]

Nvidia pays to make the games perform worse on AMD by providing stuff like "gameworks" which choke the pipeline by abusing the tesselator for example.

Add Nvidia paying developers to not use AMD features that may disadvantage Nviia products. Let remind how FutureMark were caught deliberately starving GCN cards by using Nvidia method with Time Spy.