Re: [Re] GPU reset when running the ROCm hsa runtime tests on gfx12 and next-20260701

[Bert Karwatzki]

Am Montag, dem 10.08.2026 um 18:24 +0200 schrieb Bert Karwatzki:
> Am Montag, dem 10.08.2026 um 10:15 -0400 schrieb Alex Deucher:
> > @Amber Lin
> > , @Zhang, Jesse(Jie), @Shaoyun Liu  Can you take a look at the firmware version?
> > 
> > Alex
> > 
> 
> These are the firmware version reported in sysfs:
> 
> # find /sys/class/drm/card0/device/fw_version/ -print -type f -exec cat {} \;
> /sys/class/drm/card0/device/fw_version/
> /sys/class/drm/card0/device/fw_version/pfp_fw_version
> 0x00000c12
> /sys/class/drm/card0/device/fw_version/vcn_fw_version
> 0x0910c01a
> /sys/class/drm/card0/device/fw_version/mec_fw_version
> 0x00000d0c
> /sys/class/drm/card0/device/fw_version/mes_fw_version
> 0x0000008b
> /sys/class/drm/card0/device/fw_version/dmcub_fw_version
> 0x0a000c00
> /sys/class/drm/card0/device/fw_version/me_fw_version
> 0x00000bc2
> /sys/class/drm/card0/device/fw_version/asd_fw_version
> 0x21000110
> /sys/class/drm/card0/device/fw_version/sdma2_fw_version
> 0x00798e96
> /sys/class/drm/card0/device/fw_version/sdma_fw_version
> 0x00798e96
> /sys/class/drm/card0/device/fw_version/smc_fw_version
> 0x00664600
> /sys/class/drm/card0/device/fw_version/sos_fw_version
> 0x003b0f0d
> /sys/class/drm/card0/device/fw_version/imu_fw_version
> 0x0c2c2500
> /sys/class/drm/card0/device/fw_version/rlc_fw_version
> 0x00bde160
> /sys/class/drm/card0/device/fw_version/mes_kiq_fw_version
> 0x0000008b
> 
> In addition dmesg reports a problem with the mes version:
> 
> [    4.353074] [    T489] amdgpu 0000:03:00.0: MES firmware reports incorrect version in ucode binary (0x1 vs 0x8b)
> 
> Both these result use the firmware package shipped by debian sid:
> ii  firmware-amd-graphics         20260622-1      all     Binary firmware for AMD/ATI graphics and NPU chips
> 
> The earlier result of mes_sched = 0x76 is probably from the old firmware (debian stable version 20250410), but
> the gpu resets also occur with the new firmware from debian sid.
> 
> Bert Karwatzki

I just tried the amdgpu firmware from the master branch of
https://gitlab.com/kernel-firmware/linux-firmware
and the result is the same (i.e. same firmware versions, same gpu reset).

At least the mes firmware binaries are identical:

bert@homer:/lib/firmware/amdgpu$ ls | grep gc_12 | grep mes | xargs md5sum
e50b66d9d8cae78ddf76ea27cca5a051  gc_12_0_0_mes1.bin
ed07c8a5a307bc882d42fe89441d1da9  gc_12_0_0_mes.bin
3284fd6f8269d020190d5e11075bbf6a  gc_12_0_0_uni_mes.bin
99c60afc75c11cb47007c40db29121d3  gc_12_0_1_mes1.bin
c1c9b01f5dd171906e7beefca32ebb9a  gc_12_0_1_mes.bin
3284fd6f8269d020190d5e11075bbf6a  gc_12_0_1_uni_mes.bin

This is the backed up firmware from  debian:
bert@homer:/lib/firmware/amdgpu.debian$ ls | grep gc_12 | grep mes | xargs md5sum
e50b66d9d8cae78ddf76ea27cca5a051  gc_12_0_0_mes1.bin
ed07c8a5a307bc882d42fe89441d1da9  gc_12_0_0_mes.bin
3284fd6f8269d020190d5e11075bbf6a  gc_12_0_0_uni_mes.bin
99c60afc75c11cb47007c40db29121d3  gc_12_0_1_mes1.bin
c1c9b01f5dd171906e7beefca32ebb9a  gc_12_0_1_mes.bin
3284fd6f8269d020190d5e11075bbf6a  gc_12_0_1_uni_mes.bin


Bert Karwatzki