Out-of-order compute dispatch missing on GFX12
While browsing the GFX12 code in amdgpu staging, I noticed the UNORD_DISPATCH
parameter set to 0 (as well as the CP_HQD_PQ_CONTROL__UNORD_DISPATCH_MASK
flag missing in the MQD manager code).
Previously (on GFX10 and GFX11), both the lack of these settings and the corresponding CP firmware caused a deadlock with multiple simultaneous compute job submissions, which was fixed by enabling out-of-order dispatch both in the kernel and firmware:
I realize this is a highly pre-emptive issue to report, given that no GFX12 hardware exists yet on the market, but I'm curious as to whether this is due to RDNA4 being different than previous generations or a bug. If it's the latter, we might avoid potential issues with mixed compute/graphics on GFX12 in the future.
Thanks a lot in advance and sorry for the trouble