A 42-patch series has been posted for the AMDGPU kernel driver and AMDKFD compute driver to add pipe reset support for GPU compute workloads. While queue reset has long existed to recover from GPU hangs, there are corner cases where it fails. Pipe reset addresses those cases, though it resets all queues on the affected pipe, requiring coordination across all components using compute queues. AMDGPU maintainer Alex Deucher notes the final patch still needs updating once new MES firmware is released, as older firmware may fall back to a full adapter reset.
170 Impressions