Intel Wires Up Dual-SIMD8 Dispatch For Mesa Drivers

[phoronix]

Phoronix: Intel Wires Up Dual-SIMD8 Dispatch For Mesa Drivers

Intel's open-source Linux drivers for OpenGL (Iris) and Vulkan (ANV) this week for Mesa 24.0 received support for dual-SIMD8 dispatch on Gen12 graphics (Tigerlake) and newer...

https://www.phoronix.com/news/Intel-Dual-SIMD8-Mesa

[coder]

It's also interesting to hear the doubled ALU vector width ... for Intel Xe2 graphics.

For sure. I've always found it interesting how much narrower Intel's SIMD was than competing GPUs.

The first step, in closing this gap, was creeping up to 16-lanes. I guess they're finally shooting for parity with other gaming GPUs, at 32 lanes.

[Khrundel]

Tiger lake was launched at September 2020 and will be discontinued within a month. That all you need to know about opensource GPU driver development.

[coder]

Tiger lake was launched at September 2020 and will be discontinued within a month. That all you need to know about opensource GPU driver development.

As explained in the article, the performance gain on TGL-gen iGPUs is dubious, which is why it's not (yet) on by default. Perhaps that explains why they didn't do it sooner.

The real motivation to do this is their next gen hardware.

...but, don't let me slow your troll, if it tickles your fancy.

[Khrundel]

As explained in the article, the performance gain on TGL-gen iGPUs is dubious, which is why it's not (yet) on by default. Perhaps that explains why they didn't do it sooner.

The real motivation to do this is their next gen hardware.

...but, don't let me slow your troll, if it tickles your fancy.

Nice try, but this doesn't hold any water. "No performance gain was observed and we need further evaluation" means they've expected performance gain on tigerlake and they're still expecting it after some issues will be fixed, meaning this patch is still new and they are not sure if these gains are achievable. So the question is same: why it took them 3 years to try to add this feature? I mean even assuming this optimization is useless for gen12, why they haven't found out this 3 years ago?

[aviallon]

Nice try, but this doesn't hold any water. "No performance gain was observed and we need further evaluation" means they've expected performance gain on tigerlake and they're still expecting it after some issues will be fixed, meaning this patch is still new and they are not sure if these gains are achievable. So the question is same: why it took them 3 years to try to add this feature? I mean even assuming this optimization is useless for gen12, why they haven't found out this 3 years ago?

Ah yes, I have the same kind of comments at work.
Writing software is HARD. Very hard.
Even if you have the hardware and everything else ready for things to work, things… will not work as you expect.

[coder]

Nice try, but this doesn't hold any water. "No performance gain was observed and we need further evaluation" means they've expected performance gain on tigerlake

You misquoted it, dumbass.

The article said:

"... no major performance changes have been observed on Gfx12 this series doesn't enable dual-SIMD8 by default yet until further performance evaluation is completed ..."

and they're still expecting it after some issues will be fixed,

Where did they specify anything about known issues or pending fixes?

meaning this patch is still new and they are not sure if these gains are achievable.

I think it means exactly what it says: they didn't observe any major changes, but are waiting for more evidence to know if it causes any regressions or provides any upside on currently-released products. Absent that data, they're playing it safe and leaving it off, by default.

So the question is same: why it took them 3 years to try to add this feature?

Probably because they know enough about the hardware that it got de-prioritized for having little upside potential on existing iGPUs.

I mean even assuming this optimization is useless for gen12, why they haven't found out this 3 years ago?

It was primarily added for the benefit Xe2, as the article stated. Geez. try actually reading articles, before posting like an ignoramus.

troll, troll, troll again.

[Khrundel]

You misquoted it, dumbass.

You are pathetic.
All bullshit you've written doesn't address only question that's matter: why it took them 3 years just to try? Why they're evaluating performance changes when hardware is almost EOL? You may call me troll as many times as you wish but that won't answer that.

I think it means exactly what it says: they didn't observe any major changes, but are waiting for more evidence to know if it causes any regressions or provides any upside on currently-released products. Absent that data, they're playing it safe and leaving it off, by default.

For someone who call other dumbass for "misquoting" you are not so careful with interpretations.
"Note that since no major performance changes have been observed on Gfx12 this series doesn't enable dual-SIMD8 by default yet until further performance evaluation is completed"

He said nothing about any regressions, would they exist, no performance gain could justify sending this into production. He plans to evaluate performance.

But, again, this doesn't matter. The only thing that really matters is this 3 years gap. Always be a 3rd grade citizen - that is the fate of linux users in case FOSS driver development would prevail.

[coder]

All bullshit you've written doesn't address only question that's matter:

It's been covered quite enough. The problem seems to be with your comprehension.

Perhaps Linux isn't the right OS for you. My advice would be to try MacOS or maybe Chrome OS.

[Khrundel]

It's been covered quite enough.

Sure. I just not good enough to notice it and you are too great to point it to me.
What was that? They knew it won't bring any performance gain so they did not implement it for 3 years but now they forgot about it and have implemented and got no noticeable gain but still plan to investigate further? An Goldilocks' situation here: they knew good enough to delay for 3 years but not enough just to not to expect any gain?

Or maybe it is a lot simplier: mesa is for obsolete hardware and no GPU vendor would try hard to keep it up to date with their product? I mean why would anyone invest in open software development unless they plan to save large amount of money?

Perhaps Linux isn't the right OS for you. My advice would be to try MacOS or maybe Chrome OS.

Yep, the last argument of any FOSS moron. "Get out of my linux, heretic"