Announcement

Collapse
No announcement yet.

AMD ROCm 3.9 Released With AOMP OpenMP Offloading Integrated

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

  • #41
    Originally posted by oleid View Post

    RX570? That should work fine, it is similar to Veerappan 's card. Must be your tumbleweed.

    EDIT: Just tested: blender/OpenCL works here on ArchLinux with RX470 and ROCm 3.7
    I will try again, but as far I understood it , the problem was / is with the opensourse drivers. That was not version 3.7 though.

    Comment


    • #42
      Originally posted by pixelshader View Post

      Great to hear but I believe there are more issues than that.

      I just tried to render the BMW GPU scene with ROCm 3.9 (re-)installed from scratch using your official PPA on stock Ubuntu 20.04 with a Vega 56 and a Ryzen 3950X active in parallel. Rendering does start but my mouse immediately starts stuttering and occasionally hangs when moving. Then after about 30 seconds, Blender crashes. During these 30 seconds, I saw the CPU pull ahead very quickly which is probably the performance issue you heard about and that has already been mentioned. The crash seems to be a separate issue however.
      Agreed... we found some hangs/crashes as well. Some are fixed but a couple are still being worked on.

      Comment


      • #43
        Originally posted by bridgman View Post

        Agreed... we found some hangs/crashes as well. Some are fixed but a couple are still being worked on.
        Sounds very good. I look forward to those issues being fixed and thanks very much for following up.

        Comment


        • #44
          Originally posted by bridgman View Post

          Agreed... we found some hangs/crashes as well. Some are fixed but a couple are still being worked on.
          Just wondering... Is there any roadmap you could share? That might make people more happy and not drive them back to CUDA for their personal compute needs.

          Personally, I'd love to get a new RDNA2 card and use it to drive tensorflow stuff. But if there is no hope to get it work
          in the following months, I'd need to look for alternatives. Possibly even DirectML, if I can't get it to work on Linux.

          Comment


          • #45
            Originally posted by oleid View Post
            Just wondering... Is there any roadmap you could share? That might make people more happy and not drive them back to CUDA for their personal compute needs.

            Personally, I'd love to get a new RDNA2 card and use it to drive tensorflow stuff. But if there is no hope to get it work
            in the following months, I'd need to look for alternatives. Possibly even DirectML, if I can't get it to work on Linux.
            Ask again in a month or two, OK ?

            Comment


            • #46
              Originally posted by bridgman View Post

              Ask again in a month or two, OK ?
              Will do, thanks!

              Comment


              • #47
                This 3.9 release actually improve things.

                XMR mining does work unlike with 3.8 and previous. (with is an improvement)
                Other loads seam to workish (I'm running pts/opencl right now). Some load seam to be performing about as expected, some other loads make the UI totally not responding and seam to get abysmal result.
                Also I did try FAH, it use to fold air for like 10 minutes then crash, now it does not even do that, the fah 22 process eat 100% of one CPU core and get stuck like that until I bazooka (-9) it.

                Also I seam to get a lot of Evicting PASID 0x801e queues (don't know from which test but well...)
                I'm going to leave the test suite running and see later.

                Comment

                Working...
                X