Announcement

Collapse
No announcement yet.

More Radeon Power Management Improvements

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • #46
    You need the drm-radeon-testing branch of the drm:
    http://git.kernel.org/?p=linux/kerne...radeon-testing

    Comment


    • #47
      Originally posted by agd5f View Post
      You need the drm-radeon-testing branch of the drm:
      http://git.kernel.org/?p=linux/kerne...radeon-testing
      It is not enough to use drm-radeon-testing kernel. Some X parts must be rebuilt as well: http://www.phoronix.com/forums/showthread.php?t=20484

      Anyway I am using drm-radeon-testing. Everything seems to be working. I can change profiles in sysfs but my GPU is still getting hot (Currently 89 degrees). I was trying to change profiles but nothing has helped. As well I see following messages in log:

      [ 1223.517620] radeon 0000:01:05.0: efd38000 reserve failed for wait
      [ 1325.118109] radeon 0000:01:05.0: efd38000 reserve failed for wait
      [ 1328.719790] radeon 0000:01:05.0: efd38000 reserve failed for wait
      [ 1433.991376] radeon 0000:01:05.0: efd38000 reserve failed for wait
      [ 1489.719272] radeon 0000:01:05.0: efd38000 reserve failed for wait
      [ 1576.114701] radeon 0000:01:05.0: efd38000 reserve failed for wait
      [ 1975.846892] radeon 0000:01:05.0: efd38000 reserve failed for wait
      [ 2022.020323] radeon 0000:01:05.0: efd38000 reserve failed for wait

      Comment


      • #48
        Originally posted by daliusd View Post
        It is not enough to use drm-radeon-testing kernel. Some X parts must be rebuilt as well: http://www.phoronix.com/forums/showthread.php?t=20484
        Huh, what makes you think so?

        Neither the DDX nor mesa do interface with the pm AFAIK, so why would it be necessary to rebuild X components?

        Comment


        • #49
          since alex's last commit ( http://git.kernel.org/?p=linux/kerne...86c041340079db )

          messages about changing clocks arent displaying in dmesg now, is there any way to check if reclocking actualy works and not failing due to - gui not idle ?

          Comment


          • #50
            I run glxgears then change profile from high to low, and doesnt see any noticable fps change

            Comment


            • #51
              Originally posted by netkas View Post
              I run glxgears then change profile from high to low, and doesnt see any noticable fps change
              http://wiki.cchtml.com/index.php/Glx...ot_a_Benchmark

              Comment


              • #52
                u really think I dont know it?

                lower core clock -> lower freq of triangle setup stage -> a bit slower rendering -> should result in little fps drop, but there was no drop.

                Comment


                • #53
                  Maybe pm isn't working at all for you, just like it isn't working for me

                  Comment


                  • #54
                    Originally posted by LiquidAcid View Post
                    Huh, what makes you think so?

                    Neither the DDX nor mesa do interface with the pm AFAIK, so why would it be necessary to rebuild X components?
                    Error message in X that drm failed (while dmesg says that drm is initialized). Now there is no message about drm failure.

                    OK. Maybe I'm just another blind man near elephant ;-)

                    Comment


                    • #55
                      Originally posted by netkas View Post
                      u really think I dont know it?

                      lower core clock -> lower freq of triangle setup stage -> a bit slower rendering -> should result in little fps drop, but there was no drop.
                      maybe your fps was cpu limited to begin with, so your fps didnt really change.
                      who knows

                      Comment


                      • #56
                        Originally posted by Pfanne View Post
                        maybe your fps was cpu limited to begin with, so your fps didnt really change.
                        who knows
                        I was right!

                        here is results with latest radeon-testing branch

                        14937 frames in 5.0 seconds = 2986.450 FPS
                        16194 frames in 5.0 seconds = 3238.794 FPS
                        16354 frames in 5.0 seconds = 3270.604 FPS
                        14908 frames in 5.0 seconds = 2981.395 FPS
                        switch from default to low
                        12667 frames in 5.0 seconds = 2533.344 FPS
                        12694 frames in 5.0 seconds = 2535.207 FPS
                        12616 frames in 5.0 seconds = 2522.880 FPS
                        12609 frames in 5.0 seconds = 2521.765 FPS
                        12641 frames in 5.0 seconds = 2528.155 FPS
                        switch from low to high
                        13927 frames in 5.0 seconds = 2785.204 FPS
                        14070 frames in 5.0 seconds = 2813.940 FPS


                        still memory reclocking results in screen flickerings.

                        and dynpm method seems to doesnt work at all, it just keeps using profile I set before.

                        Comment


                        • #57
                          gears is definitely cpu bound; it's basically a test of how fast your CPU can send command buffers to the GPU. If you force your CPU into it's highest power state you'll see a nice jump in performance.

                          Comment


                          • #58
                            sure, but I was measuring with same(highest) cpu speed all the way.

                            Comment


                            • #59
                              Experimenting with new PM

                              Hi all! I was (and still am) happy with my Debian Lenny & fglrx on my HP Compaq NC8430 notebook with X1600 Mobility, but reading about the advances I decided to give the open power management a try. First I tried Ubuntu Lucid, with overclock.net's howto's I installed a new kernel and I was really impressed, now my fan spins almost regularly at 55% which is where fglrx (and Windows) make it too, without the recent power management advances it would be at 70% or sometimes even at 80%.

                              Power consumption was not as satisfactory, though. I managed to get Lucid at 27.4 W, whereas Lenny with fglrx is at 20.5 W, still far from Windows at 13-14 W. Powertop was constantly reporting something about sound modules that I tried to google but didn't find any other solution than blacklisting them, but it still didn't lower the power consumption.

                              Then I thought about why not trying Arch Linux, where no bloat is installed. I used to have Arch on my previous laptop, but due to overheating the X1600 I had stopped using it on NC8430. I downloaded the newest install image, and to my surprise the 2.6.33 kernel's KMS didn't work on my computer (the screen got all striped, like a loading screen of Sinclair Spectrum's tape programs making it unreadable, although I could see that the screen was functioning "behind the stripes"). So I had to disable KMS already at the install phase (and get the fan to 80%!). I updated to the new kernel on my hot computer, only to see that it didn't have power profiles (maybe it would have had dynamic PM but I didn't try it). So then another kernel from radeon repo and voilą, I had PM even without Xorg! However, power consumption is only marginally better than Lucid's at 25.9 W so there's still some work ahead, although I'm not sure whether it is with the GPU or other components, and I haven't undervolted the CPU yet which I always do with windows, there's a kernel version conflict with PHC that I try to resolve next.

                              If it is of any use for the developers, I post here some readings of the temperature sensors as a benchmark. I don't remember (or even know) what they are exactly besides the last one that is the fan percentage. The numbering has two ways, TZ0-TZ5 or T1-T6, but the order below is from smallest to largest (the last being the fan):

                              Code:
                              Temp sensor  : T1 T2 T3 T4 T5 T6
                              Windows Vista: 50 44 43 39 20 55, power consumption 14 W
                              Ubuntu Lucid : 50 44 58 40 31 55, power consumption 27.4 W
                              Arch Linux   : 50 45 56 40 29 55, power consumption 25.9 W
                              (When I tested Lenny the computer hadn't "warmed up" for these numbers to be comparable, but the power consumption was 20.5 W)

                              Time permitting, I am willing to run new tests to see how close to Windows Linux's PM can get.

                              So a big thank you and congratulations to the developers, you are doing really well!

                              Comment


                              • #60
                                Hi, sorry for stupidly jumping in but I am kind of lost in the debris of this thread.

                                So, simple question: can I get the PM patch somewhere in a consistent state applicable to 2.6.35rc4 kernel, or maybe to 2.6.34?

                                I measured my (idle) system's power consumption. Using fglrx instead of KMS makes a difference of about 45W which is a lot in these hot days.

                                Gfx card: regular HD5770.

                                Comment

                                Working...
                                X