Benchmarks Of AMD's Newest Gallium3D Driver

[HokTar]

As if fglrx delivers the same features as catalyst...


Sorry, I am not a troll, but this one was shouting for such a reply.

Anyway, I suspect that having good h264 decoding possibilities and an efficient dynpm would significantly reduce the whining and the "what if" comments. Opengl and cl would still remain but we could have a good everyday driver. Are these on your short-term list? Maybe even helping out Christian Konig. Well, one can dream, I guess.

[bridgman]

???

In cases where the feature implementation is relatively OS-independent (ie where code *can* be shared across OSes) I believe the Linux and Windows versions of Catalyst *do* have the same features (eg OpenGL, OpenCL).

In cases where the implementation is OS-dependent or DRM is a factor (video decode acceleration, Eyefinity on older X version, even suspend/resume), ie where the code has to be pretty much re-implemented from scratch for Linux, the features on Linux can lag. Those features tend not to be a high priority for the workstation market though, so the model works OK.

For in-between things (where some of the code has to be re-implemented but not all, eg power management) there has been a lag but not so much these days.

[bridgman]

Anyway, I suspect that having good h264 decoding possibilities and an efficient dynpm would significantly reduce the whining and the "what if" comments. Opengl and cl would still remain but we could have a good everyday driver. Are these on your short-term list? Maybe even helping out Christian Konig. Well, one can dream, I guess.

Top priority for the AMD devs is enabling support for new GPU generations (via code and/or docs) and supporting community devs, which are the things that pretty much have to be done in house. When time permits, the devs also work on improving existing support (Alex wrote the initial EXA and Textured Video code, along with initial power management and a bunch of other things, Richard added GLSL and flow control support) but only when time permits.

Other than not being enabled by default, are you seeing problems with dynpm ? My recollection was that the "glitching" issue was largely fixed a month or two ago, and the remaining TODO task (splitting the code so that memory and engine clocks are changed in different blanking periods) could be done by anyone.

[HokTar]

Top priority for the AMD devs is enabling support for new GPU generations (via code and/or docs) and supporting community devs, which are the things that pretty much have to be done in house. When time permits, the devs also work on improving existing support (Alex wrote the initial EXA and Textured Video code, along with initial power management and a bunch of other things, Richard added GLSL and flow control support) but only when time permits.

Other than not being enabled by default, are you seeing problems with dynpm ? My recollection was that the "glitching" issue was largely fixed a month or two ago, and the remaining TODO task (splitting the code so that memory and engine clocks are changed in different blanking periods) could be done by anyone.

Whenever I tried it did not lowered the clock/voltage. Manually setting it to low works fine.

I think I have a fairly good understanding of the current methodology, i.e. you enable drivers than the community enhances it. I dared to ask that question because it seems that only the NI chips are left and the desktop APU is still far away. So my thinking was that in the next few months you might have some free time to help out here and there. Maybe I am completely wrong.

[bridgman]

Nope, you're right... although if we're going to stay "caught up" there's not going to be a *lot* of free time between generations.

[chrisr]

Can you elaborate on this glitching issue, please?

Other than not being enabled by default, are you seeing problems with dynpm ? My recollection was that the "glitching" issue was largely fixed a month or two ago...

Can you remember what was done to fix this "glitching" issue, please? E.g. was it in the kernel, the driver, or Mesa? Because I was noticing some glitchy behaviour with the 2.6.36.1 kernel and xorg-drv-ati from git recently (HD4890).

[HokTar]

Nope, you're right... although if we're going to stay "caught up" there's not going to be a *lot* of free time between generations.

Sometimes you show your admirable skill in not answering a question. Not to mention that I never said "a lot".

Plus I had the wild guess that the desktop apu will be based on evergreen or ni thus enabling it might not take much more than for ontario, i.e. a few hundred lines.
If this assumption turns out to be correct then there would be two devs with little to no work for months. So yes, maybe that would be a *lot* of free time.

Anyway, if Alex or Richard will have some free time then my question will get answered, I guess.

[bridgman]

Can you remember what was done to fix this "glitching" issue, please? E.g. was it in the kernel, the driver, or Mesa? Because I was noticing some glitchy behaviour with the 2.6.36.1 kernel and xorg-drv-ati from git recently (HD4890).

Pretty sure it was in the kernel. My guess is that you would need 2.6.37 to get it.

HokTar, not lowering clock/voltage is what the TODO task (splitting clock adjust for engine & memory into different vblank intervals) is intended to address. If there isn't enough time in a single vblank interval to set and stabilize both clocks the current code doesn't change either of them... and you can't lower the voltage unless the clocks have already been lowered and stabilized.

Sometimes you show your admirable skill in not answering a question. Not to mention that I never said "a lot".

No, but you implied "a lot" by suggesting that the devs work on things like assisting Christian's video decode work. The remaining work there is not "quick fixes", there's a fairly significant learning curve on the video algorithm side.

Plus I had the wild guess that the desktop apu will be based on evergreen or ni thus enabling it might not take much more than for ontario, i.e. a few hundred lines. If this assumption turns out to be correct then there would be two devs with little to no work for months. So yes, maybe that would be a *lot* of free time

A lot of the work Alex did for Ontario should carry across to Llano as well, but that's not where I see the time going. There are some differences between the NI SKUs which will require additional work on top of what has been done already, and we want to try and tie work on the next generation core into the proprietary development effort, which means we need to start *now* rather than a year from now.

I think there will be free time but I don't expect it to come quickly. When the free time *does* come the issues you mentioned may not be top-of-mind any more.

Doing more on the power management side will be a priority, whether we do it or the community devs do it. Dynpm may not be top priority either, those are the things we are trying to figure out now.

Anyway, if Alex or Richard will have some free time then my question will get answered, I guess.

It's a pretty safe bet that they will have some free time, but because it may not happen for a while I just can't give any kind of confident answer that they will be working on the specific things you mentioned.

[HokTar]

Fair enough. Thanks.

[bridgman]

It's been a few months since I mentioned trying to piggyback open source development on the proprietary driver design effort so the open source devs wouldn't have to "learn everything from scratch" and we didn't know exactly when we would be able to start...

... but it seems like the right time is "now"