Announcement

Collapse
No announcement yet.

Announcing radeontop, a tool for viewing the GPU usage

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

  • #61
    Originally posted by marek View Post
    It doesn't work on Bonaire. Everything is at 0%, only VRAM seems to be correct.
    The MMIO BAR changed on CIK parts. The relevant mapping code would need to be changed.

    Comment


    • #62
      Ideally we'd add a debugfs interface to expose the status registers so the app wouldn't have to mmap the MMIO bar itself.

      Comment


      • #63
        Patches welcome. Bonaire (7790/r260) is about five generations newer than my card, so I can't exactly test anything even close.

        Comment


        • #64
          Originally posted by agd5f View Post
          Ideally we'd add a debugfs interface to expose the status registers so the app wouldn't have to mmap the MMIO bar itself.
          Please no. Debugfs is for *debug* things only. Monitoring like this is for all users, and needs to be available without CONFIG_DEBUG* in kernel.

          Comment


          • #65
            /me cheers

            Comment


            • #66
              Anyone with a Bonaire, Kaveri, or Kabini, please test latest git. I have no such hw.

              Comment


              • #67
                Originally posted by curaga View Post
                Anyone with a Bonaire, Kaveri, or Kabini, please test latest git. I have no such hw.
                Hi,
                Tested on A10-7700K (Kaveri).
                Results looking good.
                Run some OpenCL samples over HSA and the tool definitely recognizes the increased load on the GPU.

                First benchmark:


                Second benchmark:

                Comment


                • #68
                  7790 on bus 2.
                  ./radeontop -b 2 exits
                  Code:
                  *** Error in `./radeontop': free(): invalid size: 0x00007fffc24121e0 ***
                  Aborted
                  Works correctly with line 78 of detect.c commented.

                  Comment


                  • #69
                    I assume latest git version?

                    namelist can only be invalid if scandir fails, which can only happen if you have no /dev/dri directory (or permission to read it). That's rather curious.

                    Please test latest git, added a check for that.

                    Comment


                    • #70
                      New version works perfectly, thanks!

                      Comment


                      • #71
                        help!!!

                        God job!
                        I am an openGL newbie and need your help.
                        svPerfGL (http://vis.lbl.gov/Research/svPerfGL/) is a Scientific Visualization OpenGL Graphics Benchmark using deprecated technique such as
                        display list. I altered the source. After stored the vertexes, normals and colors in VBOs, I found the FPS descend from 210 (use display list)
                        to 120 (did not use display list) with newest AMD catalyst. Here is the readeontop result:

                        1. display list

                        2,no display list


                        The two results are almost the same but differ in event engine and pf_rq_pending. I wonder what event engine and pf_rq_pending mean? By the way,
                        I add cf_rq_pending, pf_rq_pending, cp_coherency_busy, cp_busy all by myself, but do not know what they stand for. Can anybody give me a hand?
                        Thanks!

                        Comment


                        • #72
                          Originally posted by amdg View Post
                          God job!
                          I am an openGL newbie and need your help.
                          svPerfGL (http://vis.lbl.gov/Research/svPerfGL/) is a Scientific Visualization OpenGL Graphics Benchmark using deprecated technique such as
                          display list. I altered the source. After stored the vertexes, normals and colors in VBOs, I found the FPS descend from 210 (use display list)
                          to 120 (did not use display list) with newest AMD catalyst. Here is the readeontop result:

                          1. display list
                          Code:
                          http://d.pcs.baidu.com/thumbnail/6a7405ad18e20611c00f78d80a6fd605?fid=3542278631-250528-286301998249742&time=1409101200&sign=FDTAER-DCb740ccc5511e5e8fedcff06b081203-wmlUZHwRuW%2F8eG4UVWCmEJymRtk%3D&rt=sh&expires=2h&r=620937299&sharesign=unknown&size=c710_u500&quality=100
                          2,no display list
                          Code:
                          http://d.pcs.baidu.com/thumbnail/7dd19141ca5cba4d3d0c54e44c89756d?fid=3542278631-250528-1020966106899491&time=1409101200&sign=FDTAER-DCb740ccc5511e5e8fedcff06b081203-wZDNK5O%2BkLniBgIFsiz4FjYM5h4%3D&rt=sh&expires=2h&r=211544109&sharesign=unknown&size=c710_u500&quality=100
                          The two results are almost the same but differ in event engine and pf_rq_pending. I wonder what event engine and pf_rq_pending mean? By the way,
                          I add cf_rq_pending, pf_rq_pending, cp_coherency_busy, cp_busy all by myself, but do not know what they stand for. Can anybody give me a hand?
                          Thanks!
                          I hate to tell you this, but I think you're a forum newbie. You just tried to attach images that were really just links (that expire) to a pages with the images.

                          Originally posted by amdg View Post
                          God job!
                          Also, what the heck? I think you meant, "Good day!"
                          Last edited by profoundWHALE; 08-27-2014, 08:54 PM.

                          Comment


                          • #73
                            Has anyone tried the radeontop package from Ubuntu 14.04 ? When I install the package it doesn't work (see first image) but if I build from source it works fine. Obvious difference is that version information is "unknown" on the package install.



                            Comment


                            • #74
                              Originally posted by bridgman View Post
                              Has anyone tried the radeontop package from Ubuntu 14.04 ? When I install the package it doesn't work (see first image) but if I build from source it works fine. Obvious difference is that version information is "unknown" on the package install.



                              CI parts use a different BAR for MMIO. Older versions of the tool were not aware of that. I've been meaning to add a debugfs or ioctl interface to query these registers so you won't have to add as much asic specific info to the tools.

                              Comment


                              • #75
                                Originally posted by agd5f View Post
                                CI parts use a different BAR for MMIO. Older versions of the tool were not aware of that. I've been meaning to add a debugfs or ioctl interface to query these registers so you won't have to add as much asic specific info to the tools.
                                Whoops, you're right -- CI support went into radeontop just *after* 14.04 was baked. I saw the missing version info and was thinking "broken" not "old". Thanks !

                                Comment

                                Working...
                                X