FreeBSD Jumps Quickly On LLVM/Clang 3.2

[vertexSymphony]

They bark Sancho, a sign that we ride ...

[0xBADCODE]

What a "smart" decision. Sure, let's ditch GCC who generates better code in favor of pointless licensing crap. Let already slow OS become even less competitive. Good luck with such approach.

[pingufunkybeat]

Even if it was a 100% speed up they still fail to come remotely close to the blobs.

Nonsense.

Results were within 50% of the blob, often more. Add HiZ, new DMA stuff and recent general speedups and you're really close. User Verappan did benchmarks recently showing more than 60% on most benchmarks, but openbenchmarking is b0rked at this moment so I can't link it. At the same time, the lack of OpenMP slows down Clang code by a factor of 4 on most computers. 60% of the performance is "not remotely close", but 25% is sufficient. Nice logic

Of course, you have to be aware that there was a VRAM allocation regression exactly at the time of Michael's article, which is why some of the test results are wrong (Reaction Quake, xonotic, etc.) The regression has been resolved since then and results back to normal across the board. Well, mentioning that would be the honest way to go about it, anyway.

[deanjo]

Oh that's cute.

You pick the one benchmark taken in the middle of a regression that cut the framerate by 90% in many cases. And just before HiZ and the new DMA allocator were merged......

And the 27% speedup patch... http://www.phoronix.com/scan.php?pag...tem&px=MTI0OTI

http://phoronix.com/forums/showthrea...formance/page2

Even if it was a 100% speed up they still fail to come remotely close to the blobs.

[Awesomeness]

Clang is not under the BSD license. It is under the UoI-NCSA license, which is BSD-style, but different:

http://sources.gentoo.org/cgi-bin/vi...SA?view=markup

That's just splitting hairs. The ?Permission is hereby granted?? sentence and the last paragraph are verbatim copies from the MIT license and the actual conditions are a verbatim copy of the 3-clause BSD license.
There is absolutely no practical difference to the stock 3-clause BSD license. The slightly different wording is just stupid license proliferation.

[pingufunkybeat]

Lay down the eggnog and read

http://www.phoronix.com/scan.php?pag...t2012chr&num=1

Oh that's cute.

You pick the one benchmark taken in the middle of a regression that cut the framerate by 90% in many cases. And just before HiZ and the new DMA allocator were merged......

And the 27% speedup patch... http://www.phoronix.com/scan.php?pag...tem&px=MTI0OTI

http://phoronix.com/forums/showthread.php?75886-AMD-s-A-Sync-DMA-Code-Makes-For-Fast-Performance/page2

[deanjo]

HD6000 are still in production. 60-90% is not 1/5 of the performance like you claim.

Lay down the eggnog and read

http://www.phoronix.com/scan.php?page=article&item=amd_gallium_cat2012chr&num=1

[pingufunkybeat]

And how is the performance difference on cards that are not out of production?

HD6000 are still in production. 60-90% is not 1/5 of the performance like you claim.

How is the openCL performance? What is the power consumption and cpu use when watching HD video? I could go on but you should be already able to see the vast difference.

These are features, and yes, the open drivers are missing some features. Just like Clang is missing many crucial features like OpenMP. And binary drivers are missing some features too.

But we were talking about performance

I'm happy to have a decent free C++ compiler as an alternative, and I think that we all profit from that. But it's hard to deny that the major push behind Clang is licence-based.

[XorEaxEax]

They were stuck because their efforts were being put into more worthy efforts.

What are you talking about? They were stuck because of not accepting GPLv3, a purely licenced based decision. Being GPL-free has long been one of FreeBSD's stated goals, again a licence-based goal.

I'm not saying this is in any way wrong of them, they have their philosophy/ideology regarding software licencing which differs from that of FSF and it's as simple as that. But your attempts at presenting FreeBSD as making choices on a technical level rather than that of licence-philosophy falls rather flat.

Oh, and merry christmas everyone!

[deanjo]

This surely depends on the video driver.

The performance difference with r300g is smaller than the difference between Clang and GCC, for example. And r600g is at about 60% on recent cards, sometimes 90%.

And how is the performance difference on cards that are not out of production? How is the openCL performance? What is the power consumption and cpu use when watching HD video? I could go on but you should be already able to see the vast difference.