2013: A Good Year For Open-Source AMD?

[ChrisXY]

PM basically works by changing clocks and voltages to make power/performance tradeoffs that match the user's needs.

For r600 and earlier the driver code set clocks and voltages directly. On most of the more recent hardware generations the driver can still set clocks and voltages directly but each generation has different hardware blocks added which can set those clocks and voltages automatically with guidance from the driver. So far we have not been allowed to release info for those additional HW blocks, although as with UVD we have been working internally to change that. The difference is that we started internal discussions about PM a couple of years ago, concluded that we were not going to get quick approval, and that's why the current PM code was developed.

I think many people (like me) who say they would like better power management aren't even talking about automatically adjusting power levels to "match the user's needs". It's still that on "low" power level my HD 6550M drains more power than catalyst by default. (Last time I tried "low" it sometimes crashed so I used "mid" since then which seems to be exactly the same but at least appeared to be more stable).

What I personally think should be the priority in power management is bringing the card down to the minimum power drain possible for the "low" profile.

 % sudo su -c "echo low > /sys/class/drm/card0/device/power_profile"
 % cat /sys/kernel/debug/dri/0/radeon_pm_info
default engine clock: 600000 kHz
current engine clock: 299970 kHz
default memory clock: 800000 kHz
current memory clock: 299950 kHz
voltage: 900 mV
PCIE lanes: 16
 % sudo su -c "echo mid > /sys/class/drm/card0/device/power_profile"
 % cat /sys/kernel/debug/dri/0/radeon_pm_info
default engine clock: 600000 kHz
current engine clock: 299970 kHz
default memory clock: 800000 kHz
current memory clock: 299950 kHz
voltage: 900 mV
PCIE lanes: 16

[bridgman]

Btw, IIRC, you stated some years ago that R&D is trying to disentangle the problematic parts (DRM, etc.) as much as possible from the rest of the ASIC, but it will take some generations to take off. Is this something we will profit soon or has this strategy been dropped?

The strategy has not been dropped, but the HW development pipe is pretty long and we're not redesigning blocks *just* for this.

I expect the first change will become visible over the next 12-15 months, although I'm not sure we'll be able to tell you what specific changes we made. If all goes well you will see more HW info released and think "finally, what took them so long ?"

[entropy]

Nope. Same goes for all of the programming info... the best we can do is give rough estimates and confidence levels in cases where we have a specific list of tasks to complete before we can release, but this is more like UVD where we don't know what the outcome will be until the last minute.

As soon as we get to the point where the remaining issues can be discussed in public, we release

Thanks for answering.

Btw, IIRC, you stated some years ago that R&D is trying to disentangle the problematic parts (DRM, etc.) as much as possible from the rest of the ASIC,
but it will take some generations to take off. Is this something we will profit soon or has this strategy been dropped?

[DanL]

I brought that same thing up a few months ago, but I don't think Michael cares about people breaking the forum rules.

Actually, it was brought up years ago: http://phoronix.com/forums/showthrea...3-Ban-Qaridium
Watching him debate with bridgman and display mastery of the below techniques is entertaining once you get used to it. Q is just a part of the Phoronix scenery, like the beer references that some people can't stand.

Does every fcking thread have to end like this? Guys there is a function called "Ignore List". Use it!

Ignore List doesn't really work for more advanced trolls, especially those who post at the rate of Q and are skilled at threadjacking.

The Contrarian Troll. A sophisticated breed, Contrarian Trolls frequent boards whose predominant opinions are contrary to their own. It is important to distinguish between dissenters and actual Contrarian Trolls, however; the Contrarian is not categorized as a troll because of his or her dissenting opinions, but due to the manner in which he or she behaves:

? Contrarian Warning Sign Number One: The most important indicator of a poster's Contrarian Troll status is his constant use of subtle and not-so-subtle insults, a technique intended to make people angry. Contrarians will resist the urge to be insulting at first, but as their post count increases, they become more and more abusive of those with whom they disagree. Most often they initiate the insults in the course of what has been a civil, if heated, debate to that point.

? Contrarian Warning Sign Number Two: Constant references to the forum membership as monolithic. "You guys are all just [descriptor]." "You're a lynch mob." "You all just want to ridicule anyone who disagrees with you."

? Contrarian Warning Sign Number Three: Intellectual dishonesty. This is only a mild indicator that is not limited to trolls, but Contrarians display it to a high degree. They will lie about things they've said, pull posts out of context in a manner that changes their meanings significantly, and generally ignore any points for which they have no ready answers.

? Contrarian Warning Sign Number Four: Accusing the accusers. When confronted with their trolling, trolls immediately respond that it is the accusers who are trolls. Often the Contrarian will single out his most vocal opponent and claim that while he can respect his other opponents, this one in particular is beneath his notice.

? Contrarian Warning Sign Number Five: Attempts to condescend. The Contrarian will seek refuge in condescending remarks that repeatedly scorn his or her critics as beneath notice ? all the while continuing to respond to them.

? Contrarian Warning Sign Number Six: One distinctive mark of Contrarian Trolls is that every thread in which they dissent quickly devolves into a debate about who is trolling whom. In the course of such a debate the Contrarian will display many of the other Warning Signs mentioned above.

[bridgman]

How much secret sauce can there be in a "turn it off" command to UVD?

Don't think we were talking about turning UVD off -- I was talking about decoding video with it.

[curaga]

How much secret sauce can there be in a "turn it off" command to UVD?

[bridgman]

Can you share with us _why_ this turns out to be problematic?
I think it's obvious for the UVD blocks - but for the PM bits?

Nope. Same goes for all of the programming info... the best we can do is give rough estimates and confidence levels in cases where we have a specific list of tasks to complete before we can release, but this is more like UVD where we don't know what the outcome will be until the last minute.

As soon as we get to the point where the remaining issues can be discussed in public, we release

[disi]

Cards which work

I got myself a netbook ~1 year ago with an AMD C-50 APU and it works great from day one with the open driver.
In my laptop is an AMD 6970M and this card also works great (it got fried once, because I am an idiot and used conducting heat-paste which spread over the card). The Vendor shipped it to AMD and they replaced it...

Before that, I had HD4770 which was when the open driver just came out and there were screen corruptions all the time. Compared to that is the open driver close to perfect by now. Just see how much progress got accomplished in only ~3 year?!?

OK, I am an AMD Fanboy, because the last good Nvidia Card I possessed was a Riva TNT1 (still got a 460M for the laptop as replacement but don't use it)

[entropy]

The difference is that we started internal discussions about PM a couple of years ago, concluded that we were not going to get quick approval, and that's why the current PM code was developed.

Can you share with us _why_ this turns out to be problematic?
I think it's obvious for the UVD blocks - but for the PM bits?

[bridgman]

Does airlied means that there is not enough documentation to fully exploit the PM capabilities or i am reading something the wrong way??

It's complicated, so you're probably just reading it correctly

PM basically works by changing clocks and voltages to make power/performance tradeoffs that match the user's needs.

For r600 and earlier the driver code set clocks and voltages directly. On most of the more recent hardware generations the driver can still set clocks and voltages directly but each generation has different hardware blocks added which can set those clocks and voltages automatically with guidance from the driver. So far we have not been allowed to release info for those additional HW blocks, although as with UVD we have been working internally to change that. The difference is that we started internal discussions about PM a couple of years ago, concluded that we were not going to get quick approval, and that's why the current PM code was developed.

The current PM implementation sets clocks and voltages directly so it works on all generations of hardware. My thinking was that the code would need to be worked on anyways to provide better PM for r600 and earlier, and that work would benefit newer hardware as well although there would probably be a need for per-generation work to deal with HW quirks.

I think Dave is saying "why work on the current code if we know it's going to be thrown away in the future ?", which is fair, but (a) we *don't* know that the work will be thrown away in the future for newer chips (although the chances have been gradually improving over the last year or so), and (b) I believe the same work needs to be done for earlier parts anyways.