Announcement

Collapse
No announcement yet.

Radeon ROCm 4.0 Released With CDNA GPU Support (Instinct MI100)

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

  • #11
    Originally posted by bridgman View Post
    The ROCm stack releases use the same dkms kernel driver as the packaged drivers from amd.com with the same distro support - up to 20.04.1 at the moment.

    Are you sure that the packaged OpenGL and OpenCL drivers are being installed correctly on PopOS ? That seems more likely at first glance than Resolve not recognizing the drivers because you used --no-dkms. Maybe check the output of "glxinfo | grep render" and clinfo as a first step ?

    Does *anyone* actually use DaVinci Resolve on RHEL/CentOS (the distro that Blackmagic supports, and a primary target of our packaged drivers) ?
    Hello John. Does ROCm work with Renoir APUs? I have a laptop that has both Renoir and NVidia graphics. I use NVidia graphics to develop and test our existing HPC software relying on CUDA. I'd like to try to replace CUDA with HIP and see whether our software stack can work on AMD and Nvidia hardware using HIP. I guess HIP doesn't support running on both cards simultaneously and requires separate compilation for each platform but it is still nice to see if I can use HIP with AMD graphics only. So is it possible to make Renoir work with ROCm? Documentation explicitly mentions the lack of support of APUs but it seems like this section hasn't been updated for a while.

    Comment


    • #12
      Originally posted by Aisyk View Post
      Hi,

      I'm a little lost with theses AMD drivers, OpenCL, Clover, OpenCL 1.1, 1.2, 20...

      I have a navi card (5500XT) and i want to use DaVinciResolve with OpenCL (it doesn't launch without this support). I'm on a PopOS 20.10 distribution.
      Will theses RocM drivers can enable OpenCL for use with DaVinciResolve ?

      Proprietary OpenCL drivers don't compile with 20.10 kernels... i need to use a -nodkms option to install them but DaVinciResolve don't recognize the right version...
      I feel your pain. AMD + Liinux + OpenCL has been a train wreck of contradictory announcements and directions. If even Michael is having trouble keeping it all sorted out then you know it is bad. I'm sure you have seen it already but if not there is a mega thread on the BMD forum about installing on Linux.



      Originally posted by walterav View Post

      Are you using AMDGPU Pro drivers aswell for instance the legacy OpenGL component or full Opensource graphics stack including ROCM? Since D.R. 16.1 and higher required the closed AMDGPU Pro Legacy OpenGL driver, before "h33p" found a solution.

      BTW are D.R. 17beta's working for you?
      I am running an RX580 and 16 still works while 17 doesn't. It will be a while yet as I am pretty sure that Apple is throwing money at BMD to get the M1 port to a gold state. After that they will focus on Windows and then after that Linux.

      Originally posted by bridgman View Post
      Does *anyone* actually use DaVinci Resolve on RHEL/CentOS (the distro that Blackmagic supports, and a primary target of our packaged drivers) ?
      My main desktop is Fedora and then I dual boot to Centos 8 solely for resolve. I am running an older version of AMDGPU-pro drivers there. It took me a month or so to finally find a combination that would work with resolve and I will only touch it again at gun point. I would prefer to run ROCm there but my understanding is that it stopped working with resolve with the ROCm 2.0 release.

      Would I prefer to just run DR out of my Fedora desktop and avoid the PITA of the dual boot and would I prefer to use a more modern desktop distro as opposed to a server distro? Yup, sure would. I suspect there are some large production houses that run the BMD supplied image but among the majority of the base almost no one does.

      Comment


      • #13
        I hope Michael is able to lay hands on one of those MI100 cards at some point as I am curious to see how it's compute performance stacks up against the gaming cards.

        Comment


        • #14
          Originally posted by MadeUpName View Post
          I hope Michael is able to lay hands on one of those MI100 cards at some point as I am curious to see how it's compute performance stacks up against the gaming cards.
          Doubtful, last time I had any pro/workstation cards from AMD was pre-GCN FirePro...
          Michael Larabel
          https://www.michaellarabel.com/

          Comment


          • #15
            Originally posted by Solid State Brain View Post
            If only I could manage to install it on OpenSUSE Tumbleweed. Tried many times for OpenCL support in Blender with my RX480; no success so far.
            This won't work cause of the mixing of llvm versions.

            Comment


            • #16
              Originally posted by Rakot View Post
              Hello John. Does ROCm work with Renoir APUs? I have a laptop that has both Renoir and NVidia graphics. I use NVidia graphics to develop and test our existing HPC software relying on CUDA. I'd like to try to replace CUDA with HIP and see whether our software stack can work on AMD and Nvidia hardware using HIP. I guess HIP doesn't support running on both cards simultaneously and requires separate compilation for each platform but it is still nice to see if I can use HIP with AMD graphics only. So is it possible to make Renoir work with ROCm? Documentation explicitly mentions the lack of support of APUs but it seems like this section hasn't been updated for a while.
              Right now I think the answer is "no" but AFAIK Renoir is the closest to being supportable. Checking...
              Test signature

              Comment


              • #17
                Originally posted by bridgman View Post
                Does *anyone* actually use DaVinci Resolve on RHEL/CentOS (the distro that Blackmagic supports, and a primary target of our packaged drivers) ?
                Not CentOS but Fedora on my case. ROCN OpenCL is literally broken when attempting to run on Raven Ridge Ryzen 2500u when it was used for Davinci Resolve. AMDGPU-Pro version is unofficially supported for that APU but the GPU part is listed as unknown thus preventing Davinci Resolve to boot.
                Honestly, I failed to understand why APU especially the mobile version remained neglected on both open source and official AMD driver.

                Comment


                • #18
                  RDNA / RDNA2: OpenCL works somehow, but with bad performance, the rest: no not really..
                  APUs / most consumer Vega Cards: Not officially supported, but they might work

                  Obviously AMD does not really have the bandwidth / resources to support consumer hardware / even care.. The Accelerators are GCN (Vega 20 for example) and CDNA based.. So currently CDNA is Data Center / Accelerator only currently and probably it will stay that way.. At least if games don´t suddently start to use bf16 and such, we won´t see consumer cards based on CDNA IMHO.

                  Just a guess:
                  RDNA: Has Raytracing units, but lacks Matrix multiplication accelerators and bf16 format
                  CDNA: Lacks raytracing units, but has matrix multiplication accelerator and bf16 format + probably more cores and a lot higher memory Bandwidth

                  Comment


                  • #19
                    Originally posted by finalzone View Post
                    but the GPU part is listed as unknown thus preventing Davinci Resolve to boot.
                    That's the first I have heard of this - are you talking about the string that appears in lspci, or the renderer string, or something else ?

                    Originally posted by finalzone View Post
                    Honestly, I failed to understand why APU especially the mobile version remained neglected on both open source and official AMD driver.
                    I don't think it is neglected on the open source side, at least for GPU. There were some CPU / OS issues with the first Raven parts but I think we're past that now.

                    Not sure what you mean by "official AMD driver" since upstream is the primary official AMD driver - guessing you're talking about the prebuilt driver packages on amd-com, either the all-open built from an upstream fork or the hybrid workstation drivers with open source kernel/libdrm/multimedia and closed source OpenGL/Vulkan ?

                    The business unit responsible for APUs has historically not felt that there was much demand for workstation/compute support on APUs, but that is starting to change.

                    Originally posted by Spacefish View Post
                    RDNA / RDNA2: OpenCL works somehow, but with bad performance, the rest: no not really..
                    APUs / most consumer Vega Cards: Not officially supported, but they might work
                    Again this is the opposite of my understanding. We made a big push on OpenCL / RDNA for the 20.45 release, moving it from a PAL back end to the ROCm back end and fixing a lot of correctness & performance issues. If you're talking about the big Blender slowdown, we were finally able to root-cause and fix that. RDNA2 is in a bit better shape than RDNA1 in 20.45 but we are working through the deltas for the next release.

                    Originally posted by Spacefish View Post
                    Obviously AMD does not really have the bandwidth / resources to support consumer hardware / even care.. The Accelerators are GCN (Vega 20 for example) and CDNA based.. So currently CDNA is Data Center / Accelerator only currently and probably it will stay that way.. At least if games don´t suddently start to use bf16 and such, we won´t see consumer cards based on CDNA IMHO.

                    Just a guess:
                    RDNA: Has Raytracing units, but lacks Matrix multiplication accelerators and bf16 format
                    CDNA: Lacks raytracing units, but has matrix multiplication accelerator and bf16 format + probably more cores and a lot higher memory Bandwidth
                    Pretty close:

                    - RDNA has some specialized instructions for ML (eg dot product operations, more in RDNA2) but does not have matrix acceleration.

                    - CDNA has no graphics pipeline at all (so no ray tracing units) but has matrix multiplication, bf16 and a lot more cores, along with very high bandwidth
                    Test signature

                    Comment


                    • #20
                      Originally posted by bridgman View Post

                      Right now I think the answer is "no" but AFAIK Renoir is the closest to being supportable. Checking...

                      I would really appreciate if you can clarify it.

                      On a side note, AMD won a number of DOE awards for building next-gen exascale capable clusters that we (scientists) are going to use. Good software support is a must have for us. AMD is partnering with Kokkos team but it is not enough if we cannot even test already available hardware. On Nvidia side, despite terrible open source support and a number of problems in general like recent "GPL condom" incident, we can develop and test our software stack on both mobile and HPC video cards. This is quite handy.

                      Comment

                      Working...
                      X