Re: RDNA 3: Kernel has requested more VGPRs than are available
Hi Cory,
On 2024-07-24 19:09, Cordell Bloor wrote:
> Hello,
>
> I noticed a gfx1100 failure on the CI [1]:
>
> 608s :0:rocdevice.cpp :2690: 1613980217 us: [pid:953 tid:0x7f00d4e006c0]
> Callback: Queue 0x7effce100000 aborting with error :
> HSA_STATUS_ERROR_OUT_OF_REGISTERS: Kernel has requested more VGPRs than
> are available on this agent code: 0x2d
> 608s Aborted
>
> The incorrect VGPR count on Navi 31 is a bug in firmware-amd-graphics
> and should be fixed in 20240610, which is now available on unstable and
> testing. This test was running on 'explorer'. What firmware version was
> it using?
'explorer' is a bookworm host with firmware from 20230210 from bookworm,
and kernel 6.7 from bookworm-backports.
When running in a VM, the upstream-binaries test driver emits this
information (and dmesg) as an artifact, see [1] for example. Adding a
workaround or fallback for podman runners to export this information is
on my TODO list.
Best,
Christian
[1]: https://ci.rocm.debian.net/packages/r/rocsparse/unstable/amd64+gfx1030/
Reply to: