PackageCompiler fails to create app for MadNLPGPU + ExaModels (CUDSS linear solver)

[KSepetanc]

PackageCompiler fails on both LTS (1.10.8) and current stable Julia (1.11.3) to build a package with MadNLPGPU + ExaModels dependencies (cuDSS linear solver and --precompile=all). There are no other dependencies. I've created MWE and described the issue on Discourse. GPU optimization app would be great contribution to Julia features. Can we try to fix this? Maybe even make a test?

[KSepetanc]

Not sure if this will help, but PrecompileTools also has issue compiling MadNLPGPU with ExaModels. err command of @compile_workload shows:

<Regular MadNLP log of successful run is above>

EXIT: Optimal Solution Found (tol = 1.0e-08).
LLVM ERROR: Cannot select: intrinsic %llvm.nvvm.membar.sys

[172740] signal 6 (-6): Aborted
in expression starting at none:0
Allocations: 425799162 (Pool: 425779641; Big: 19521); GC: 8

[KSepetanc]

Smaller MWE is that PackageCompiler does not work when only CUDA.jl is included. Example code is:

create_sysimage([“CUDA”], sysimage_path=“image.so” , sysimage_build_args=--compile=all)

More detailed report is on Discourse link above.

[yuwenchen95]

Not sure if this will help, but PrecompileTools also has issue compiling MadNLPGPU with ExaModels. err command of @compile_workload shows:

<Regular MadNLP log of successful run is above>

EXIT: Optimal Solution Found (tol = 1.0e-08).
LLVM ERROR: Cannot select: intrinsic %llvm.nvvm.membar.sys

[172740] signal 6 (-6): Aborted
in expression starting at none:0
Allocations: 425799162 (Pool: 425779641; Big: 19521); GC: 8

I also encounter the same issue for precomiling the CuClarabel package:

PackageCompiler: compiling incremental system imageLLVM ERROR: Cannot select: intrinsic %llvm.nvvm.membar.sys

Any suggestion to avoid this issue?

Btw, I have set incremental = false, but why I still need to compile the system incrementally?

[KSepetanc]

I appreciate that you joined the conversation. It seems that everything CUDA.jl related fails the compile (presumably with the error somewhere in the GPUCompiler.jl). I wish to raise the importance of this if possible. It is needed feature.

[KristofferC]

Seems like JuliaGPU/GPUCompiler.jl#611 is where the conversation should be. PackageCompiler doesn't do anything out of the ordinary here and there is likely nothing that can be done here to fix it.