Bug 68224

Created attachment 84473 [details] [review]
Possible Fix

Can you try this mesa patch.  If it doesn't fix the crash, can you re-post the shader dump with the patch applied.

(In reply to comment #1)
> Created attachment 84473 [details] [review] [review]
> Possible Fix
> 
> Can you try this mesa patch.  If it doesn't fix the crash, can you re-post
> the shader dump with the patch applied.

With the patch applied it still fail

Created attachment 84475 [details]
Shader dump with patch applied

Created attachment 84543 [details] [review]
Patch #2

Created attachment 84544 [details] [review]
LLVM Patch #3

If you apply all three of these patches, does it fix the crash?

(In reply to comment #5)
> Created attachment 84544 [details] [review] [review]
> LLVM Patch #3
> 
> If you apply all three of these patches, does it fix the crash?

It's better, now it crash before the intro start (so latter), new shader dump attached

Created attachment 84551 [details]
Shader dump with 3 patches applied

Same problem with Brütal Legend (after the intro):

Buddha.bin.x86: AMDGPUInstrInfo.cpp:109: virtual void llvm::AMDGPUInstrInfo::storeRegToStackSlot(llvm::MachineBasicBlock&, llvm::MachineBasicBlock::iterator, unsigned int, bool, int, const llvm::TargetRegisterClass*, const llvm::TargetRegisterInfo*) const: Assertion `!"Not Implemented"' failed.
Stack dump:
0.      Running pass 'Function Pass Manager' on module 'tgsi'.
1.      Running pass 'Greedy Register Allocator' on function '@main'

Created attachment 84880 [details] [review]
SGPR register spilling patch

Can you test this patch?  It should fix the crash, but it may hang your GPU.

(In reply to comment #9)
> Created attachment 84880 [details] [review] [review]
> SGPR register spilling patch
> 
> Can you test this patch?  It should fix the crash, but it may hang your GPU.

It's not really better; instead of the assert, it gives a GPU lockup

(In reply to comment #9)
> .addImm(Lane);
	
This results in the lane number being encoded verbatim in the VSRC1 instruction field, which I don't think is correct?

I tried adding 0x80 to make it an inline literal, but Sanctuary still locks up.

Another potential gotcha I noticed in the ISA documentation is that the V_READ/WRITELANE instructions don't respect the EXEC mask.

Created attachment 85372 [details] [review]
SGPR register spilling patch v2

Can you try this v2 patch?  It fixes the bug Michel found plus another one.

(In reply to comment #12)
> Created attachment 85372 [details] [review] [review]
> SGPR register spilling patch v2
> 
> Can you try this v2 patch?  It fixes the bug Michel found plus another one.

Same result as patch v1, GPU lockup instead of llvm assert; Sanctuary demo gives also a GPU lockup

(In reply to comment #13)
> (In reply to comment #12)
> > Created attachment 85372 [details] [review] [review] [review]
> > SGPR register spilling patch v2
> > 
> > Can you try this v2 patch?  It fixes the bug Michel found plus another one.
> 
> Same result as patch v1, GPU lockup instead of llvm assert; Sanctuary demo
> gives also a GPU lockup

Sanctuary doesn't lockup for me with this patch, but the only thing visible is the torch.  Everything else is black.  What settings are you using with Sanctuary?

(In reply to comment #14)
> (In reply to comment #13)
> > (In reply to comment #12)
> > > Created attachment 85372 [details] [review] [review] [review] [review]
> > > SGPR register spilling patch v2
> > > 
> > > Can you try this v2 patch?  It fixes the bug Michel found plus another one.
> > 
> > Same result as patch v1, GPU lockup instead of llvm assert; Sanctuary demo
> > gives also a GPU lockup
> 
> Sanctuary doesn't lockup for me with this patch, but the only thing visible
> is the torch.  Everything else is black.  What settings are you using with
> Sanctuary?

It's locking up just at the end of the loading screen, so it's the default settings

(In reply to comment #15)
> (In reply to comment #14)
> > (In reply to comment #13)
> > > (In reply to comment #12)
> > > > Created attachment 85372 [details] [review] [review] [review] [review] [review]
> > > > SGPR register spilling patch v2
> > > > 
> > > > Can you try this v2 patch?  It fixes the bug Michel found plus another one.
> > > 
> > > Same result as patch v1, GPU lockup instead of llvm assert; Sanctuary demo
> > > gives also a GPU lockup
> > 
> > Sanctuary doesn't lockup for me with this patch, but the only thing visible
> > is the torch.  Everything else is black.  What settings are you using with
> > Sanctuary?
> 
> It's locking up just at the end of the loading screen, so it's the default
> settings

Are you using libtxc-dxtn and/or do you have floating-point textures enabled?

(In reply to comment #16)
> 
> Are you using libtxc-dxtn and/or do you have floating-point textures enabled?

I have both enabled, (lib32-)libtxc-dxtn and floating-point textures.

(In reply to comment #12)
>  - Use correct register flags, so we don't overwrite the low dword
>    when restoring multi-dword registers.

This doesn't seem to result in any difference for the code generated for the first Sanctuary shader which spills registers.

Now Sanctuary works fine with llvm-3.4svn r194070 / mesa-git 86cdff5 @ PITCAIN

Brutal Legend / Serious Sam 3 are still problematic

With Mesa 10.1.0-devel (git-0601598) & LLVM 3.4 (svn-r195266) still quits in the same moment but log is a little bit different:
Sam3: AMDGPUInstrInfo.cpp:113: virtual void llvm::AMDGPUInstrInfo::storeRegToStackSlot(llvm::MachineBasicBlock&, llvm::MachineBasicBlock::iterator, unsigned int, bool, int, const llvm::TargetRegisterClass*, const llvm::TargetRegisterInfo*) const: Assertion `!"Not Implemented"' failed.

I don't try Brutal Legent, but Stalker game(by the wine) has the same issue and log changes from AMDGPUInstrInfo.cpp:113 to AMDGPUInstrInfo.cpp:109 after update mesa/llvm too.

Created attachment 89616 [details] [review]
Work Around

Hopefully this patch will fix the crash, but it will likely result in some things being rendered incorrectly.  Can you test it and let me know if it makes the game playable.

Created attachment 89631 [details]
dmesg with fix (stalker)

Hi, i can't get log from SS3 because GPU crashes and can't get tty or ssh.
Wine doesn't give me any usefull logs but gpu was restarted succesfully and i have dmesg output. 

I applied fix on top on latest llvm(3.5). Should i use 3.4 branch instead?

I'll try redirect steam's logs to non-temporary filesystem and maybe that gives me some info.

Created attachment 89640 [details]
steam's log (SS3)

My PC hung. This is log from the game - nothing exciting as far as i know.
When i find out anything useful i will call you.

(In reply to comment #23)
> Created attachment 89640 [details]
> steam's log (SS3)
> 
> My PC hung. This is log from the game - nothing exciting as far as i know.
> When i find out anything useful i will call you.

If you could get a dump with the environment variable

R600_DEBUG=ps,vs

That would help.

Created attachment 89659 [details]
log with R600_DEBUG=ps,vs

llvm with Tom's fix

Created attachment 89671 [details] [review]
Debugging Patch

Do you still have a GPU hang with this patch?  Can you post the output of R600_DEBUG=ps,vs with this patch?

Created attachment 89680 [details]
v2 log with R600_DEBUG=ps,vs

yes it still hung.

Created attachment 89719 [details]
v3 log with R600_DEBUG=ps,vs

i have to downgrade libdrm to 2.4.48 because it gives me randomly gpu crash.

may be related to bug 71859

Created attachment 89806 [details] [review]
Possible Fix

Can you try this patch?  If you get a lockup or a hang, please post the output of R600_DEBUG=ps,vs

Created attachment 89825 [details]
v4 log with R600_DEBUG=ps,pv

it's not hung anymore, but new log:

Sam3: SIInstrInfo.cpp:98: virtual void llvm::SIInstrInfo::copyPhysReg(llvm::MachineBasicBlock&, llvm::MachineBasicBlock::iterator, llvm::DebugLoc, unsigned int, unsigned int, bool) const: Assertion `AMDGPU::SReg_32RegClass.contains(SrcReg)' failed.

BTW stalker & xonotic works perfectly with this fix. 
(32bit 10.1 git-ddc77c5 + llvm 3.5 r195722+fix)

Stalker? When did they release stalker for linux???

We have wine don't you forget;) Now [thanks Tom] it works with dynamic lighting etc.

Created attachment 89853 [details] [review]
Patch 1

Created attachment 89854 [details] [review]
Patch 2

Can you try these two patches?

(In reply to comment #35)
> Created attachment 89854 [details] [review] [review]
> Patch 2
> 
> Can you try these two patches?

You should apply them both together.

yes, it works. Some textures flicker, but i don't know how it work before. 
I don't see visual regressions in unigine-*, xonotic, lightmarks.
I check driver performance with/without patches later. 

thanks a lot Tom

performance is fine, slow but not slower than before :)

The fix was committed to the LLVM tree: r195880

Created attachment 89954 [details]
screenshot

It works but it's unplayable, 25 fps with shit graphic settings and 5 fps with Ultra graphic settings.