Failed to translate LLVM code to SPIR-V on PVC

[michel2323]

I get this error when using oneAPI.jl on Oceananigans.jl on PVC. I tried both oneAPI@v2.0.1 and v2.0.3 with the new SPIR-V translator. This is the output with v2.0.3. The code is a mapreduce implementing a sum! on a custom type (Field). I've also attached the file that is recommended for the bug report. I can try to reduce this into an MWE if this is not helpful.

jl_POTmPmRCT2.zip

InvalidInstruction: Can't translate llvm instruction:
 Global variable cannot have Function storage class. Consider setting a proper address space.
 Original LLVM value:
@_j_const1 = private unnamed_addr constant i64 0, align 8
ERROR: LoadError: Failed to translate LLVM code to SPIR-V.
If you think this is a bug, please file an issue and attach /var/tmp/pbs.4904394.aurora-pbs-0001.hostmgmt.cm.aurora.alcf.anl.gov/jl_POTmPmRCT2.bc.
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:35
  [2] macro expansion
    @ /lus/flare/projects/Julia/mschanen/julia_depot_path/packages/GPUCompiler/fi1Sa/src/spirv.jl:153 [inlined]
  [3] mcgen(job::GPUCompiler.CompilerJob{GPUCompiler.SPIRVCompilerTarget, oneAPI.oneAPICompilerParams}, mod::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler /lus/flare/projects/Julia/mschanen/julia_depot_path/packages/GPUCompiler/fi1Sa/src/utils.jl:134
  [4] macro expansion
    @ /lus/flare/projects/Julia/mschanen/julia_depot_path/packages/Tracy/GcShf/src/tracepoint.jl:158 [inlined]
  [5] macro expansion
    @ /lus/flare/projects/Julia/mschanen/julia_depot_path/packages/GPUCompiler/fi1Sa/src/driver.jl:404 [inlined]

[maleadt]

Coincidentally, @vchuravy also ran into this recently with KA.jl+PoCL (OpenCL.jl). I haven't looked into the sources of this constant, but this will probably need a change in GPUCompiler.jl a la https://github.com/JuliaGPU/GPUCompiler.jl/blob/bdf390a561498a2c8c6b7de8f71206953515901e/src/metal.jl#L339-L416

[foglienimatteo]

I'm working on a PVC too with Julia 1.10.9 and oneAPI v2.0.3. I get the same InvalidBitWidth error as in #55 :

InvalidBitWidth: Invalid bit width in input: 7
ERROR: LoadError: Failed to translate LLVM code to SPIR-V.
If you think this is a bug, please file an issue and attach /hppfs/scratch/08/di75tom/tmp/jl_WLC9oVGBCx.bc.
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:35
  [2] macro expansion
    @ ~/.julia/packages/GPUCompiler/sy6sk/src/spirv.jl:153 [inlined]
  [3] mcgen(job::GPUCompiler.CompilerJob{GPUCompiler.SPIRVCompilerTarget, oneAPI.oneAPICompilerParams}, mod::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/sy6sk/src/utils.jl:134
  [4] macro expansion
    @ ~/.julia/packages/Tracy/GcShf/src/tracepoint.jl:158 [inlined]
  [5] macro expansion
    @ ~/.julia/packages/GPUCompiler/sy6sk/src/driver.jl:404 [inlined]
  [6] macro expansion
    @ ~/.julia/packages/Tracy/GcShf/src/tracepoint.jl:158 [inlined]
  [7] macro expansion
    @ ~/.julia/packages/GPUCompiler/sy6sk/src/driver.jl:401 [inlined]
  [8] emit_asm(job::GPUCompiler.CompilerJob, ir::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/sy6sk/src/utils.jl:116

I attach the file jl_WLC9oVGBCx.bc as recommended by the logs.
I'm not understanding what is the underlying issue. Is there a temporary workaround?

jl_WLC9oVGBCx.bc.tar.gz

[maleadt]

That is a separate issue.

[vchuravy]

So this issue is coming from SPIRV-LLVM-Translator and is expected https://github.com/KhronosGroup/SPIRV-LLVM-Translator/blob/221039c032ebce082cf505d2cda85b6fada1adc6/test/negative/GlobalVarNoAddrspace.ll#L12

In my case, it is https://github.com/JuliaGPU/GPUCompiler.jl/blob/772a78c7dac90064d9fa9d9aead8ac199f9773a6/src/irgen.jl#L217 that is not setting the correct address-space.