(In reply to Christoph Haag from comment #2)
> [47303.471211] radeon 0000:01:00.0:   VM_CONTEXT1_PROTECTION_FAULT_ADDR  
> 0x0FF03E4E

The fact that bits 32-39 of the faulting addresses are FF indicates incorrect
shader code generation resulting in those address bits being clobbered.

Can you generate an apitrace which reproduces the Packet0 error and/or GPUVM
faults?