Announcement

Collapse
No announcement yet.

AMD AOMP 11.5 Released For OpenMP Offloading To Radeon GPUs

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

  • AMD AOMP 11.5 Released For OpenMP Offloading To Radeon GPUs

    Phoronix: AMD AOMP 11.5 Released For OpenMP Offloading To Radeon GPUs

    Released on Wednesday was AOMP 11.5 as the latest version of the AMD/ROCm compiler based off LLVM Clang and focused on OpenMP offloading to Radeon GPUs...

    http://www.phoronix.com/scan.php?pag...-11.5-Released

  • #2
    does it work on APUs as well?

    Comment


    • #3
      @Michael: How do you feel about benchmarking this AOMP beast? There are a few OpenMP in the test suite. Might be interesting.

      FYI: I recently had quite good experiences using ROCm/tensorflow on ClearLinux and ArchLinux using their official docker container (https://hub.docker.com/r/rocm/tensorflow). For my R470, I had to use the official dkms driver on the host system, though. For more recent cards, however, the mainline kernel driver could work.

      Comment


      • #4
        Originally posted by szymon_g View Post
        does it work on APUs as well?
        It would seem there is no _official_ support for ROCm on APUs, however, I found this page: https://bruhnspace.com/en/bruhnspace-rocm-for-amd-apus/

        They say:
        THE APUS ARE ONLY SUPPORT IN THE UPSTREAM LINUX KERNEL DRIVER.
        You could try to use the ROCm/tensorflow container I linked above and try to run it on your APU. Maybe it already works.

        Comment


        • #5
          Originally posted by oleid View Post
          @Michael: How do you feel about benchmarking this AOMP beast? There are a few OpenMP in the test suite. Might be interesting.
          Those OpenMP tests do not work with offloading.

          Comment


          • #6
            Originally posted by Setif View Post

            Those OpenMP tests do not work with offloading.
            Oh, why not?

            Comment


            • #7
              Originally posted by oleid View Post

              Oh, why not?
              OpenMP offloading was initially introduced in OpenMP-4.0 and it needs special directives, for example
              Code:
              #pragma omp parallel for ...
              works only with multicore CPUs .
              To offload code to GPU you have to write
              Code:
              #pragma omp target teams distribute parallel for ...
              and move explicitly data from host to device or from device to host by
              Code:
              #pragma target data ...

              Comment


              • #8
                Ah, that's unfortunate. I didn't follow OpenMP to closely. Last time I used it was v3. Thanks for the explanation!

                Comment


                • #9
                  Originally posted by oleid View Post

                  It would seem there is no _official_ support for ROCm on APUs, however, I found this page: https://bruhnspace.com/en/bruhnspace-rocm-for-amd-apus/

                  They say:

                  You could try to use the ROCm/tensorflow container I linked above and try to run it on your APU. Maybe it already works.
                  That page is interesting and frankly I have not seen it before. However it would be best if AMD got to work on updating RCOm to fully support APU's and even Navi (I'm not even sure of the state of Navi support). It has been a very long time since Raven Ridge and now we have Renoir. It would be nice to see AMD putting more effort into this. I do realize though that they have been seriously engaged in bug hunting in the GPU drivers so maybe we can give them a pass for a few more months.

                  Comment


                  • #10
                    Originally posted by wizard69 View Post

                    That page is interesting and frankly I have not seen it before. However it would be best if AMD got to work on updating RCOm to fully support APU's and even Navi (I'm not even sure of the state of Navi support). It has been a very long time since Raven Ridge and now we have Renoir. It would be nice to see AMD putting more effort into this. I do realize though that they have been seriously engaged in bug hunting in the GPU drivers so maybe we can give them a pass for a few more months.
                    I've kinda given up on decent open source compute support

                    That's my 6630M, M395X and Raven 3400G all not had support, there was partial support with Clover but I could never get any real life apps working with it

                    Comment

                    Working...
                    X