[Steam Deck] Blender Crashes when selecting HIP compute

[Mhowser]

Hello,
Not sure if this is the right place to submit this bug report, but since this project is the closest I have ever been to getting the VanGogh GPU working for Blender, I think it fits.

What I have is Blender 3.4.1, downloaded straight off the Official website. my Steam Deck is running stock SteamOS 3.4.4 (latest stable) with some tweaks to pacman for getting the required packages. rocm-hip-sdk and rocm-opencl-sdk + all deps were installed just fine. I was missing the hsa-amd-aqlprofile-bin package, fixed that. Now I am stuck here, on this crash when I click on the HIP render device button in Blender settings:

$ ./blender
Read prefs: /home/deck/.config/blender/3.4/config/userpref.blend
/run/user/1000/gvfs/ non-existent directory
blender: /build/hsa-rocr/src/ROCR-Runtime-rocm-5.4.1/src/image/addrlib/src/gfx10/gfx10addrlib.cpp:1071: virtual rocr::Addr::ChipFamily rocr::Addr::V2::Gfx10Lib::HwlConvertChipFamily(unsigned int, unsigned int): Assertion `false' failed.
Aborted (core dumped)

Here is some relevant information that might help:
rocminfo.txt
system-infoCLI.txt
glxinfo.txt
vulkaninfo.txt

Output of inxi -G:

$ ./inxi -G
Graphics:
  Device-1: AMD VanGogh [AMD Custom GPU 0405] driver: amdgpu v: kernel
  Display: x11 server: X.Org v: 21.1.4 with: Xwayland v: 21.1.99 driver: X:
    loaded: amdgpu unloaded: modesetting gpu: amdgpu resolution:
    1: 1920x1080~60Hz 2: 1280x800~60Hz
  OpenGL: renderer: AMD Custom GPU 0405 (vangogh LLVM 14.0.6 DRM 3.45
    5.13.0-valve36-1-neptune) v: 4.6 Mesa 22.2.0 (git-17e5312102)

I tried $ ./blender --debug-gpu-force-workarounds, $ ./blender --env-system-python HSA_OVERRIDE_GFX_VERSION=10.3.0 and $ HSA_OVERRIDE_GFX_VERSION=10.3.0 ./blender with no spectacular results, just the same crash.

Please let me know if you require some more information and how to properly get it for you.

[tpkessler]

Hey! This is a good place for this issue. Since all involved packages are also in [community-testing] you can also use the official Arch Linux bug tracker. The problem in your case is that mobile GPUs are not officially supported. So some calls (like rocminfo and there-like) may work but more complex code like that generated by blender may fail. See https://github.com/rocm-arch/rocm-arch#package-x-does-not-work

[Mhowser]

looking at https://github.com/RadeonOpenCompute/ROCR-Runtime/blob/ac0384d6256b4910509c257a30d1a13d01b8a5ee/src/image/addrlib/src/gfx10/gfx10addrlib.cpp it seems like the VanGogh GPU is supported? (I dont know anything about C++) and the problem might be that the method to get information about the GPU is not sending the correct value. This is my uneducated guess.

 case FAMILY_VGH:
            if (ASICREV_IS_VANGOGH(chipRevision))
            {
                m_settings.supportRbPlus   = 1;
                m_settings.dccUnsup3DSwDis = 0;
            }
            else
            {
                ADDR_ASSERT(!"Unknown chip revision");
            }
            break;

[tpkessler]

As long as the architecture is not listed on the official list, there's not much we can do. You can try, but it may fail unexpectedly.

[syberphunk]

Fascinating, since the Van Gogh (VN) APU has the Cadence Vision Q6 and Q5 cores.

[splatchoot]

Any update on the subject? I can actually select the GPU and click on the render button in the viewport but I end up with this error "HIP hipcc compiler not found. Install HIP toolkit in default location."