RADV Vulkan Driver Switches To The RTLD Linker

[phoronix]

Phoronix: RADV Vulkan Driver Switches To The RTLD Linker

A big set of patches landed this morning in RADV Git for Mesa 19.2...

http://www.phoronix.com/scan.php?page=news_item&px=RADV-Vulkan-RTLD-Linker

[BNieuwenhuizen]

More like recent LLVM git broke LDS usage if you did not use RTLD (which is critical for tessellation and compute shaders). So the "improvements" are really that it fixes an LLVM introduced regression.

[M@yeulC]

I have been hitting a few hangs with both radeonsi and RADV using my R9 Fury, how should I go about debugging those, given that the system does not recover, thus gives no logs?

Should I try to address the recovery issue first?

Example hang: https://bugs.freedesktop.org/show_bug.cgi?id=111020

[andrei_me]

BNieuwenhuizen this new feature seems cool and important, can you explain what it is or what it does like the other ELF thing a couple months ago?

Google tells me that RTLD means "Runtime Loader" but I don't know how it means in the driver context

[smitty3268]

BNieuwenhuizen this new feature seems cool and important, can you explain what it is or what it does like the other ELF thing a couple months ago?

Google tells me that RTLD means "Runtime Loader" but I don't know how it means in the driver context

This is the same thing as the recent radeonsi change, just for radv now.

[shmerl]

More like recent LLVM git broke LDS usage if you did not use RTLD (which is critical for tessellation and compute shaders). So the "improvements" are really that it fixes an LLVM introduced regression.

Ah, thanks for the heads up! I was bitten by the above when building Mesa master with recent llvm 9 snapshot. I'll retry again now.

[Grogan]

Same here, I was at first blaming Vulkan in my mesa-git build, but it turned out to be LLVM 9. Suspecting miscompilation of shaders, on a hunch, I went back to LLVM 8 and rebuilt. Since 19.1.0 was released, I've been using release tags rather than master though.

Some games using Vulkan (native and DXVK) were getting what seemed to be corrupted lighting and shadows. Using LLVM 8 fixed that. I'm happy to see it being addressed now.

[Brisse]

I have been hitting a few hangs with both radeonsi and RADV using my R9 Fury, how should I go about debugging those, given that the system does not recover, thus gives no logs?

Should I try to address the recovery issue first?

Example hang: https://bugs.freedesktop.org/show_bug.cgi?id=111020

I used to encounter hangs like that on my Fury. Usually it came down to LLVM issues. Haven't had them in a while though, but I haven't tried the game you mentioned either so it could be this game is triggering some specific issue.

[M@yeulC]

I used to encounter hangs like that on my Fury. Usually it came down to LLVM issues. Haven't had them in a while though, but I haven't tried the game you mentioned either so it could be this game is triggering some specific issue.

Well, I also have hangs in No Man's Sky (OpenGL game), and had one in Ryse: son of rome (DXVK). That specific game is also using DXVK, so maybe I should report directly to DXVK devs? But it's also the only game where I can reliably reproduce the issue.

[Brisse]

Well, I also have hangs in No Man's Sky (OpenGL game), and had one in Ryse: son of rome (DXVK). That specific game is also using DXVK, so maybe I should report directly to DXVK devs? But it's also the only game where I can reliably reproduce the issue.

For Vulkan applications, did you try both RADV and AMDVLK? Sometimes one will cause hangs but not the other. Could be useful information while debugging.