3 Replies Latest reply on Apr 29, 2009 7:31 PM by ccmwtf

    Severe Problems with CAL and Catalyst 9.4 drivers

    Gipsel
      does not run under XP systems

      I've got feedback from quite some users of my application claiming the CAL interface of the Cat9.4 just does not run correctly under WindowsXP (32 and 64bit versions). After some seconds to some minutes of intense GPU load the driver/CAL crashes (with or without a VPU recover event). That was not a problem with Catalyst versions 8.12 to 9.2. It appears to work under Vista though.

      I'm aware that Cat 9.3 introduced a bug to the CAL libraries causing a quite similar behaviour under WinXP that was corrected with a hotfix. The release notes of the Cat 9.4 claim that bug is also fixed with the newest driver release, but it appears that this may be not the full truth.

      Has anybody experienced similar problems with the Catalyst 9.4 driver?

      Edit: Just for clarification, the kernels are not running several seconds or minutes. The kernels are designed to run less than 50ms each. But I call them in some kind of interleaved manner (using different tasks and different contexts) to get maximum GPU utilization. If you do just that, it crashes quite often with a VPU recover (should not do it either). So I went forward and synchronized the kernel calls with a mutex. The next kernel (from a different thread) is only called when the kernel before has really completed. That worked up to Cat 9.2 but stopped afterwards. It runs for some time but then randomly crashes the driver and simply do not return from a kernel call.

        • Severe Problems with CAL and Catalyst 9.4 drivers
          Crashtest

          Yes - as you know I got this problem with NT5.2 x64 (both srv03 and xp) with 9.4 and MilkyWay@Home

           

          Mfg/Regards

          RK - P3D

          • Severe Problems with CAL and Catalyst 9.4 drivers
            Gipsel

             

            Originally posted by: Gipsel

            Edit: Just for clarification, the kernels are not running several seconds or minutes. The kernels are designed to run less than 50ms each. But I call them in some kind of interleaved manner (using different tasks and different contexts) to get maximum GPU utilization. If you do just that, it crashes quite often with a VPU recover (should not do it either). So I went forward and synchronized the kernel calls with a mutex. The next kernel (from a different thread) is only called when the kernel before has really completed. That worked up to Cat 9.2 but stopped afterwards. It runs for some time but then randomly crashes the driver and simply do not return from a kernel call.



            Strange thing is it runs under Vista no matter which driver version you throw at it. Generally, the compatibility of the Catalyst drivers slowly went downhill from the 8.12. But the 9.3 and 9.4 are simply not working at all here under XP. Even if I don't interleave kernel calls from several threads (GPU load down to 70% or so) it crashes with Cat 9.3 and 9.4. It only takes a bit longer, so it may run a minute or two with a slow HD3000 series card. A HD4890 still crashes within less than a minute.

            Something has to be wrong with the drivers Looking to Folding@home, they also appear to have some issues with driver stability.

              • Severe Problems with CAL and Catalyst 9.4 drivers
                ccmwtf

                Any luck getting around this?  I'm seeing same thing (maybe a minute of invocations before VPU recover) after getting my kernel debugged and running (4850, 1.4, XP).  I cut it way back and it still fails, and even floyd_warshall -i 10000 dies.  I wanted to try it when the card wasn't driving video, but my BIOS won't let me run onboard video with the PCIE card plugged in,  so I'm done until I upgrade to new mobo or switch to NV.