RDNA3 on Linux

[gnusenpai]

It sounds like using a particular driver on Windows allows RDNA3 to work. However, on Linux, none of the 3 Vulkan drivers are currently capable of generating images.

Here are the drivers tested and the results:

• RADV (Mesa 23.0.0-devel git-e584447aed): This one segfaults when trying to compile the vae2base_19dec_fp16_tuned_vulkan vmfb.
• AMDVLK (2022.Q4.4) / AMDGPU-PRO (22.40): These drivers compile the vmfbs without errors, but the generated images are black.

I had to employ a few workarounds to get this far, so I'll list them here in case they might be the cause:

1. Torchvision was not installed by default. I manually installed it with the command I posted on the relevant issue here: https://github.com/nod-ai/SHARK/issues/703#issuecomment-1365307370

2. The Vulkan target triple is not detected correctly on RADV, AMDVLK, or AMDGPU-PRO drivers. I assume this is because the device names are different depending on the platform/driver used. On Windows, I think this is AMD Radeon RX 7900 XTX, but on Linux, it is AMD Radeon Graphics (RADV GFX1100) on RADV and AMD Radeon Graphics on AMDVLK and AMDGPU-PRO. I used the following patch to workaround this:

   diff --git a/shark/iree_utils/vulkan_utils.py b/shark/iree_utils/vulkan_utils.py
   index 9c73eaa..cd10f9e 100644
   --- a/shark/iree_utils/vulkan_utils.py
   +++ b/shark/iree_utils/vulkan_utils.py
   @@ -86,10 +86,8 @@ def get_vulkan_target_triple(device_name):
        triple = f"pascal-gtx1080-{system_os}"
   
    # Amd Targets
   -    elif all(x in device_name for x in ("AMD", "7900")):
   -        triple = f"rdna3-7900-{system_os}"
    elif any(x in device_name for x in ("AMD", "Radeon")):
   -        triple = f"rdna2-unknown-{system_os}"
   +        triple = f"rdna3-7900-{system_os}"
    else:
        triple = None
    return triple

System specs: OS: Gentoo Linux Kernel: 6.2.0-rc1 Python: 3.10.9 GPU: AMD Radeon RX 7900 XTX

[gnusenpai]

Well this is interesting. I figured I'd try using AMDGPU-PRO without the 2nd workaround just to see what would happen, and sure enough, it actually works. I guess this means that something with the tuned models doesn't quite work under Linux, so there is still something a bit wrong.

[powderluv]

Nice to see another fellow Gentoo user.

RADV has known issues. If your vulkaninfo | grep deviceName doesn't show 7900 xtx then it runs as rdna2 so it is slower. You can pass target-triple-flag in command line for now.