AMDVLK Driver Updated With New Extension, Better Geometry Shader Support

[phoronix]

Phoronix: AMDVLK Driver Updated With New Extension, Better Geometry Shader Support

The AMD developers working on their official, cross-platform Vulkan driver have carried out another weekly update to their AMDVLK public source tree...

http://www.phoronix.com/scan.php?page=news_item&px=AMDVLK-On-Chip-GS-New-Ext

[humbug]

Nice, the race is on.

I'll be working on some fresh AMDVLK vs. RADV benchmark numbers shortly

Any chance you can throw in one or two equivalent MSRP Nvidia GPUs also on their proprietary driver?
e.g. gtx 1060 to compare to rx 580
or GTX 1080 to compare to Vega64

[duby229]

Also it would be nice to see one or two low end GPU's in the mix so that scalibility can be observed.

[perpetually high]

Wasn't able to get the last build to compile, wasn't able to get this one to compile.

/opt/VulkanSDK/AMDVLK/drivers/llvm/lib/Analysis/DomPrinter.cpp: In constructor ‘{anonymous}::DomOnlyViewer::DomOnlyViewer()’:
/opt/VulkanSDK/AMDVLK/drivers/llvm/lib/Analysis/DomPrinter.cpp:298:1: internal compiler error: Segmentation fault
 }
 ^
Please submit a full bug report,
with preprocessed source if appropriate.
See <file:///usr/share/doc/gcc-7/README.Bugs> for instructions.
llvm/lib/Analysis/CMakeFiles/LLVMAnalysis.dir/build.make:710: recipe for target 'llvm/lib/Analysis/CMakeFiles/LLVMAnalysis.dir/DomPrinter.cpp.o' failed
make[2]: *** [llvm/lib/Analysis/CMakeFiles/LLVMAnalysis.dir/DomPrinter.cpp.o] Error 1
CMakeFiles/Makefile2:3693: recipe for target 'llvm/lib/Analysis/CMakeFiles/LLVMAnalysis.dir/all' failed
make[1]: *** [llvm/lib/Analysis/CMakeFiles/LLVMAnalysis.dir/all] Error 2
make[1]: *** Waiting for unfinished jobs....

Any ideas?

edit: Ubuntu 17.10 and gcc 7.2.0

[smitty3268]

Is Vega fully supported yet?

Docs still say the following, but maybe that's the mysterious better geometry shader support?

Dawn of War III may crash during gameplay on Radeon™ RX Vega Series due to geometry shader is not yet supported in LLPC

[lu_tze]

Is Vega fully supported yet?

Docs still say the following, but maybe that's the mysterious better geometry shader support?

From the AMDVLK readme:

Note: Pipeline compiler support for geometry and tessellation shaders is not fully implemented for the Radeon™ RX Vega Series. APU support is limited. These issues will be addressed in upcoming releases.

So probably not yet in full.

[lu_tze]

Wasn't able to get the last build to compile, wasn't able to get this one to compile.

/opt/VulkanSDK/AMDVLK/drivers/llvm/lib/Analysis/DomPrinter.cpp: In constructor ‘{anonymous}::DomOnlyViewer::DomOnlyViewer()’:
/opt/VulkanSDK/AMDVLK/drivers/llvm/lib/Analysis/DomPrinter.cpp:298:1: internal compiler error: Segmentation fault
}
^
Please submit a full bug report,
with preprocessed source if appropriate.
See <file:///usr/share/doc/gcc-7/README.Bugs> for instructions.
llvm/lib/Analysis/CMakeFiles/LLVMAnalysis.dir/build.make:710: recipe for target 'llvm/lib/Analysis/CMakeFiles/LLVMAnalysis.dir/DomPrinter.cpp.o' failed
make[2]: *** [llvm/lib/Analysis/CMakeFiles/LLVMAnalysis.dir/DomPrinter.cpp.o] Error 1
CMakeFiles/Makefile2:3693: recipe for target 'llvm/lib/Analysis/CMakeFiles/LLVMAnalysis.dir/all' failed
make[1]: *** [llvm/lib/Analysis/CMakeFiles/LLVMAnalysis.dir/all] Error 2
make[1]: *** Waiting for unfinished jobs....

Any ideas?

edit: Ubuntu 17.10 and gcc 7.2.0

How are you building it?

I have similar problem with

make -j

and LLVM 6+, but in my case, it exhausts all the memory and wakes up OOM killer, which then randomly kills stuff like X11. Building just with

make

finishes successfully, just takes an eternity.

[perpetually high]

How are you building it?

I have similar problem with

make -j

and LLVM 6+, but in my case, it exhausts all the memory and wakes up OOM killer, which then randomly kills stuff like X11. Building just with

make

finishes successfully, just takes an eternity.

Nice call, man. Thanks.

make by itself worked where make -j4 didn't, like you mentioned.

[haagch]

A debug build with LTO requires insane amounts of memory, over 3 GB (!) per make job.

It gets a bit better when you do a release build with -DCMAKE_BUILD_TYPE=Release

[lu_tze]

A debug build with LTO requires insane amounts of memory, over 3 GB (!) per make job.

It gets a bit better when you do a release build with -DCMAKE_BUILD_TYPE=Release

It was able to exhaust 32 GB RAM + 16 GB swap just with -j 2 and Release build. The interesting thing is, that "-G Ninja" and then building with ninja works fine, with all the cores you throw at it.

Unfortunately, ROCm's hcc has similar problem and there building with ninja doesn't work.