Ptxas: unresolved extern function __spirv_AtomicLoad(long long const*...)

[12VIP]

Issue (ptxas error) observed while building sample application(after dpct migration)

In this case :
During compilation phase object file TestFluxBC.cpp.dp.o is generated then ,during linking phase :
command :
dpcpp_compiler/bin/clang++ -fsycl -fsycl-targets=nvptx64-nvidia-cuda -o TestFluxBC tests/TestFluxBC.cpp.dp.o liblbpm-wia.a -L/usr/lib/x86_64-linux-gnu/openmpi/lib -lmpi_cxx -lmpi

Error Reported:
ptxas fatal : Unresolved extern function '_Z18__spirv_AtomicLoadPKxN5__spv5Scope4FlagENS_19MemorySemanticsMask4FlagE'
llvm-foreach:
ptxas fatal : Unresolved extern function '_Z18__spirv_AtomicLoadPKxN5__spv5Scope4FlagENS_19MemorySemanticsMask4FlagE'
llvm-foreach:
clang++: error: ptxas command failed with exit code 255 (use -v to see invocation)
clang version 16.0.0 (/netbatch/donb3373736_00/runDir/jenkins-dir/workspace/Tools_SH/SYCLOS_Nightly/RHEL/llvm.src/clang 2359d94)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: sycl_nightly_20221203/dpcpp_compiler/bin
clang++: note: diagnostic msg: Error generating preprocessed source(s).

Function :
c++filt '_Z18__spirv_AtomicLoadPKxN5__spv5Scope4FlagENS_19MemorySemanticsMask4FlagE'
__spirv_AtomicLoad(long long const*, __spv::Scope::Flag, long long const::MemorySemanticsMask::Flag)

• OS: Linux(Ubuntu)
• DPC++ version: 16.0.0(2359d94)
• Dependencies : openmpi

Will try to check on the reproducer steps(as lots of steps prerequisite) .Want to check if this is some known issue .

[mmoadeli]

Hi @12VIP thanks for reporting this issue. We will look into this and get back to you shortly. Would be helpful if you could share the related code to reproduce the issue.

[dm-vodopyanov]

Quick look on the error message: the reason the compilation failed is that CUDA doesn't support SPIR-V.

[bader]

Quick look on the error message: the reason the compilation failed is that CUDA doesn't support SPIR-V.

CUDA doesn't use SPIR-V, so I don't think this is the right guess. I suppose it has something to do with the libclc implementation, which is supposed to implement this function. The reason why this function definition is missing might be that long long type is not supported by OpenCL. @AerialMantis, could you assign someone to investigate, please?

[12VIP]

files.zip : Attached cpp file & headers if required.

[zjin-lcf]

@12VIP
Do your files contain DPC++ codes ?

[12VIP]

@zjin-lcf , Original code : https://github.com/OPM/LBPM/blob/master/tests/TestFluxBC.cpp
Attached earlier(files.zip with dpct) : TestFluxBC.cpp.dp.cpp .Both files are same only ,no dpc++ code as such .

[jchlanda]

@12VIP I wasn't able to build your example (as attached in files.zip) due to broken dependencies, it would be great if you could provide us with a minimal SYCL example that exhibits the same issue. Recently we have merged in fixes to re-mangler, which might be of help in that case (as long long would likely be requiring re-mangline), further more I suggest you try this PR: #7391 as it introduces more support for generic address space atomics.

[jchlanda]

In fact I've just had a look in the generated libclc files (from that PR I've mentioned above) and indeed it contains the function in question:

remangled-l32-signed_char.libspirv-nvptx64--nvidiacl.ll

479632:declare dso_local i64 @_Z18__spirv_AtomicLoadPKxN5__spv5Scope4FlagENS_19MemorySemanticsMask4FlagE(i64* noundef, i32 noundef, i32 noundef) local_unnamed_addr #50