An LLVM-Based Intel GPU 3D Driver Won't Come Soon

[mattst88]

What a vacuous pile of propaganda.

Oh! Fantastic! I'm really glad to hear that.

I'm sure you can help smooth out the problems with R600g?

[Marc Driftmeyer]

Please stop with the conspiracy theory.

LLVM is a nightmare to use in your project. See all of the difficulties with using it in r600g.

What a vacuous pile of propaganda.

[mattst88]

I don't think it's a conspiracy theory. All you have to do is go back a year or two and look up the mailing list messages, and you'll see that a lot of the Intel devs that talked about LLVM didn't seem real wild about it.

No, but this part is:

It sounds to me like Ian (or whoever) has managed to put a halt to this project for now. We'll see if it ever starts up again...

[curaga]

lol @ Bridgman

[bridgman]

The real question IMO is what hints the beer and glass pictured in the article (I think it's called a Kwak glass, originally created to facilitate drinking and driving ) are supposed to be suggesting, other than "people talk about compilers while drinking beer".

[smitty3268]

Please stop with the conspiracy theory.

LLVM is a nightmare to use in your project. See all of the difficulties with using it in r600g.

I don't think it's a conspiracy theory. All you have to do is go back a year or two and look up the mailing list messages, and you'll see that a lot of the Intel devs that talked about LLVM didn't seem real wild about it.

In particular, it would mean having to throw away a bunch of the new code they've just written for the last couple years, and that's always hard to do when you are invested in a project.

That's why i was so surprised when the initial Intel/LLVM news came out, and much less surprised by this news.

It would be like Intel suddenly announcing they were going to use Gallium instead of their custom classic driver architecture.

[bridgman]

Yep... my understanding from XDC in Chicago was that the Intel devs had :

- lots of awareness of potential and benefits of LLVM
- lots of awareness of downsides and challenges too
- general feeling that switching over immediately would be really painful (and it was)

We agreed at the conference that radeon would push ahead with LLVM and see how it went... and if it went smoothly then the other driver teams would probably follow. We knew that LLVM would be a painful transition, but we wanted to be able to make progress on both graphics (using TGSI) and OpenCL (using LLVM IR) at the same time and so going straight to LLVM seemed worth the risk.

Hopefully the work we've been doing will make it less painful for others to adopt LLVM in the future, but it's probably fair to say that not a month passed without me lying awake at night at least once wondering if we should have started with a "simple" TGSI-to-ISA translator for SI instead... even after Tim took over managing the open source graphics team and I moved to HSA.

For what it's worth, I don't think the challenges are all that LLVM-specific -- my experience has been that changing compiler frameworks is usually painful, particularly when you go from a simpler framework to a more complex one or a custom framework to a generic one (ie it's always painful because nobody ever goes the other way).

On a positive note, Tom's hair hasn't turned white yet, so it couldn't have been *that* much of a nightmare.

[mattst88]

I get the sense that some people, such as Ian, were strongly opposed to the idea from the very start, which is why i was surprised to hear that Intel actually had plans to move to LLVM.

It sounds to me like Ian (or whoever) has managed to put a halt to this project for now. We'll see if it ever starts up again...

Please stop with the conspiracy theory.

LLVM is a nightmare to use in your project. See all of the difficulties with using it in r600g.

[bjwbell]

GPGPU Wording Hints...

Michael's wording on the status of GPGU computing hints that there will be something official to report later.

My guesses are (1) an official opensource OpenCL GPU project (2) a closed source port of the Windows OpenCL GPU driver or (3) continue what we have now which is no support.

I think (3) is unlikely and I'm hoping for (1). I know it would take less engineering manpower to do (2).

A continuation of beignet would be great.

Disclaimer: I work at Intel but not in OTC and have no inside knowledge.

[smitty3268]

I get the sense that some people, such as Ian, were strongly opposed to the idea from the very start, which is why i was surprised to hear that Intel actually had plans to move to LLVM.

It sounds to me like Ian (or whoever) has managed to put a halt to this project for now. We'll see if it ever starts up again...