The Linux 4.19 AMDGPU + Mesa 18.3 Performance From The Radeon HD 7950 To RX Vega 64

[M@yeulC]

I succeeded in getting some info after crash using your recommended command "journalctl -k -b -1":

Sep 30 21:33:00 ckiws02 kernel: amdgpu 0000:01:00.0: GPU fault detected: 146 0x000e880c
Sep 30 21:33:00 ckiws02 kernel: amdgpu 0000:01:00.0: VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x00000000
Sep 30 21:33:00 ckiws02 kernel: amdgpu 0000:01:00.0: VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x0E08800C
Sep 30 21:33:00 ckiws02 kernel: amdgpu 0000:01:00.0: VM fault (0x0c, vmid 7) at page 0, read from '' (0x00000000) (136)
Sep 30 21:33:22 ckiws02 kernel: amdgpu 0000:01:00.0: GPU fault detected: 146 0x000e880c
Sep 30 21:33:22 ckiws02 kernel: amdgpu 0000:01:00.0: VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x00000000
Sep 30 21:33:22 ckiws02 kernel: amdgpu 0000:01:00.0: VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x0E08800C
Sep 30 21:33:22 ckiws02 kernel: amdgpu 0000:01:00.0: VM fault (0x0c, vmid 7) at page 0, read from '' (0x00000000) (136)
...... and many more

Until now I didn't find useful comments resp. this error. Maybe you have a hint how to proceed.

Reg.
Christof

Sorry to answer late. Unfortunately, I am not too knowledgeable about this. I would ask in #amdgpu on freenode (with a link to the complete log, and mention VM_CONTEXT1_PROTECTION_FAULT), and wait for someone to answer me, to be frank.

You might also have a bit of luck with enabling GPU recovery:

(/sys/kernel/debug/dri/0/amdgpu_gpu_recover), see https://www.phoronix.com/forums/foru...54#post1036954

Though I'm not sure whether your freeze is a complete system freeze, or just GPU? (sounds like the latter). Can you still ssh in? You can also boot with amdgpu.gpu_recovery=1 to enable it automatically.

Though I wasn't personally very lucky with this (but haven't had crashes/hangs in a long time).

[Bonnom]

I think around a year ago I had problems with weird screen artifacts (some screen lines jittering up and down) on my r9 290 on 4k. Reported that bug to on the amd site and like a month later it was fixed. Pretty cool that they still fix driver issues for the older cards!

[chris61]

Damn, lost half my post :/

Thanks for the precision Michael, hopefully that's just a temporary regression on git.

chris61, seems like OpenSuse only has llvm5, you could try to upgrade that, it often fixes crashes (maybe have a look at the flatpak xdg runtime?). Other than that, you might be able to obtain a log trough ssh, pstore, or a serial port. Just make sure that you dont't have it with journalctl -k -b -1 before



I had to look that up: https://www.reddit.com/r/Amd/comment...haders/e0uv6bd
Basically a rasterization shader. But I still don't know what NGG means (I've looked, though).

I succeeded in getting some info after crash using your recommended command "journalctl -k -b -1":

Sep 30 21:33:00 ckiws02 kernel: amdgpu 0000:01:00.0: GPU fault detected: 146 0x000e880c
Sep 30 21:33:00 ckiws02 kernel: amdgpu 0000:01:00.0: VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x00000000
Sep 30 21:33:00 ckiws02 kernel: amdgpu 0000:01:00.0: VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x0E08800C
Sep 30 21:33:00 ckiws02 kernel: amdgpu 0000:01:00.0: VM fault (0x0c, vmid 7) at page 0, read from '' (0x00000000) (136)
Sep 30 21:33:22 ckiws02 kernel: amdgpu 0000:01:00.0: GPU fault detected: 146 0x000e880c
Sep 30 21:33:22 ckiws02 kernel: amdgpu 0000:01:00.0: VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x00000000
Sep 30 21:33:22 ckiws02 kernel: amdgpu 0000:01:00.0: VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x0E08800C
Sep 30 21:33:22 ckiws02 kernel: amdgpu 0000:01:00.0: VM fault (0x0c, vmid 7) at page 0, read from '' (0x00000000) (136)
...... and many more

Until now I didn't find useful comments resp. this error. Maybe you have a hint how to proceed.

Reg.
Christof

[DanglingPointer]

...Although the Hawaii doesn't have full DX12 support

From https://wccftech.com/amd-confirms-gc...-120-gcn-1112/
AMD DirectX 12 GCN Support:

Radeon HD 7000 series	GCN 1.0	DX12, feature level 11_1
Radeon HD 7790	GCN 1.1	DX12, feature level 12_0
Radeon R7 260 (X)	GCN 1.1	DX12, feature level 12_0
Radeon R9 270 (X)	GCN 1.0	DX12, feature level 11_1
Radeon R9 280 (X)	GCN 1.0	DX12, feature level 11_1
Radeon R9 285	GCN 1.2	DX12, feature level 12_0
Radeon R9 290 (X)	GCN 1.1	DX12, feature level 12_0

[Kraut]

NGG on Vega is broken in silicon or facing the mining rush bullshit they decided to only implement it if that would have crashed early one. Now they would just damage future GPU sales if the marked gets flushed with ex. mining Vegas that got a significant rasterization speedup via "magic" drivers.

[kiffmet]

NGG (Next-Gen Geometry) was meant to increase the geometry throughput of the GPU 3x-4x through Primitive Shaders with special intrinsics (not the PS used right now as these are just basic merged vertex and geometry shaders) in tandem with some other architectural features and shader compiler trickery. It would also keep the CUs from being underutilized (one of the big problems with GCN). My hope was that support for this could be baked into DXVK, but well...

Maybe someone can hack together a driverside implementation for Vega, once this is available for Navi.

https://lists.freedesktop.org/archiv...st/025320.html

I wonder if bridgman or agd5f are allowed tell something about this whole situation?

[M@yeulC]

Damn, lost half my post :/

Thanks for the precision Michael, hopefully that's just a temporary regression on git.

chris61, seems like OpenSuse only has llvm5, you could try to upgrade that, it often fixes crashes (maybe have a look at the flatpak xdg runtime?). Other than that, you might be able to obtain a log trough ssh, pstore, or a serial port. Just make sure that you dont't have it with journalctl -k -b -1 before

It's a shame that NGG won't be coming to Vega anymore. I would have loved to see that card mature with big leaps in performance, although the current performance is also not that bad. More important then that would be a seamless and hassle-free Freesync implementation on Linux with the open drivers though.

I had to look that up: https://www.reddit.com/r/Amd/comment...haders/e0uv6bd
Basically a rasterization shader. But I still don't know what NGG means (I've looked, though).

[M@yeulC]

Oops, forgot to mention it in the article. For that particular configuration, RoTR would hang with Fury.

Thanks! Well, it ran fine last time, let's hope it's just a temporary regression due to bleeding edge code.

[...] OpenSUSE Leap 15.0/Mesa (last try: Kernel 4.19 RC2/Mesa 18.2 [not only Xplane11] crashes but the complete system crashes[...]
Can anybody give me a hint how to proceed, are there possibilities to get a log even if the machine is freezing?

You don't get a journal trough journalctl -k -b -1 ? Another possibility would be trough pstore/serial, or even launching dmesg -w trough ssh at the same time (in my experience, it spits out quite a bit of things before hanging the system, when a GPU fault occurs).
Second "wild guess": It seems like the llvm version on OpenSuse is a bit outdated (5.0), did you try to upgrade that? You could run it out of a flatpak, for instance (I think the xdg runtime includes some more recent llvm libs).

It's a shame that NGG won't be coming to Vega anymore. I would have loved to see that card mature with big leaps in performance, although the current performance is also not that bad. More important then that would be a seamless and hassle-free Freesync implementation on Linux with the open drivers though.

I had to look that up: https://www.reddit.com/r/Amd/comment...haders/e0uv6bd
Basically a rasterization shader. But I still don't know what NGG means (primitive shader? I've looked, though).

[cybertraveler]

Look at the results for the Radeon HD 7950 running DOTA 2. The average, the min & the max are all pretty much identical for both tests. It looks like there is some kind of bottleneck. My [uninformed] guess is that the bottleneck could be the relatively slow speed of the GPU memory. Both tests were done at 3840 x 2160. That's a lot of pixels being read, processed and written on that old hardware.

http://gpuboss.com/gpus/Radeon-RX-480-vs-Radeon-HD-7950

Michael - what do you think?

[chris61]

I'm using the DVD version. Maybe that's the difference. I have to test next week.
Thanks for your hint.
Reg.
Christof