Division for Int128 not defined on GPU

Specific Domains GPU
,
1

Here is a MWE.

julia> using CUDA

julia> x = widen.(CuArray(rand(Int64, 10)))
10-element CuArray{Int128, 1}:
  7760362279116003288
  1887985077088633303
 -4343813946050421943
 -3266615630613073703
  2140177075751050950
 -8479154113856376281
 -2068873100572326709
 -8353273203968219407
  6541739557653674531
  8433965827099195653

julia> .÷(x, x)
ERROR: LLVM error: Undefined external symbol "__divti3"
Stacktrace:
  [1] handle_error(reason::Cstring)
    @ LLVM ~/.julia/packages/LLVM/1GCWB/src/core/context.jl:105
  [2] macro expansion
    @ ~/.julia/packages/LLVM/1GCWB/src/util.jl:85 [inlined]
  [3] LLVMTargetMachineEmitToMemoryBuffer
    @ ~/.julia/packages/LLVM/1GCWB/lib/libLLVM_h.jl:705 [inlined]
  [4] emit(tm::LLVM.TargetMachine, mod::LLVM.Module, filetype::LLVM.API.LLVMCodeGenFileType)
    @ LLVM ~/.julia/packages/LLVM/1GCWB/src/targetmachine.jl:44
  [5] mcgen(job::GPUCompiler.CompilerJob, mod::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/DwnNd/src/mcgen.jl:74
  [6] macro expansion
    @ ~/.julia/packages/TimerOutputs/PZq45/src/TimerOutput.jl:226 [inlined]
  [7] macro expansion
    @ ~/.julia/packages/GPUCompiler/DwnNd/src/driver.jl:331 [inlined]
  [8] macro expansion
    @ ~/.julia/packages/TimerOutputs/PZq45/src/TimerOutput.jl:226 [inlined]
  [9] macro expansion
    @ ~/.julia/packages/GPUCompiler/DwnNd/src/driver.jl:328 [inlined]
 [10] emit_asm(job::GPUCompiler.CompilerJob, ir::LLVM.Module; strip::Bool, validate::Bool, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/DwnNd/src/utils.jl:62
 [11] cufunction_compile(job::GPUCompiler.CompilerJob)
    @ CUDA ~/.julia/packages/CUDA/mVgLI/src/compiler/execution.jl:313
 [12] cached_compilation(cache::Dict{UInt64, Any}, job::GPUCompiler.CompilerJob, compiler::typeof(CUDA.cufunction_compile), linker::typeof(CUDA.cufunction_link))
    @ GPUCompiler ~/.julia/packages/GPUCompiler/DwnNd/src/cache.jl:87
 [13] cufunction(f::GPUArrays.var"#broadcast_kernel#16", tt::Type{Tuple{CUDA.CuKernelContext, CuDeviceVector{Int128, 1}, Base.Broadcast.Broadcasted{Nothing, Tuple{Base.OneTo{Int64}}, typeof(div), Tuple{Base.Broadcast.Extruded{CuDeviceVector{Int128, 1}, Tuple{Bool}, Tuple{Int64}}, Base.Broadcast.Extruded{CuDeviceVector{Int128, 1}, Tuple{Bool}, Tuple{Int64}}}}, Int64}}; name::Nothing, kwargs::Base.Iterators.Pairs{Union{}, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ CUDA ~/.julia/packages/CUDA/mVgLI/src/compiler/execution.jl:288
 [14] cufunction
    @ ~/.julia/packages/CUDA/mVgLI/src/compiler/execution.jl:282 [inlined]
 [15] macro expansion
    @ ~/.julia/packages/CUDA/mVgLI/src/compiler/execution.jl:102 [inlined]
 [16] #launch_heuristic#261
    @ ~/.julia/packages/CUDA/mVgLI/src/gpuarrays.jl:17 [inlined]
 [17] launch_heuristic
    @ ~/.julia/packages/CUDA/mVgLI/src/gpuarrays.jl:17 [inlined]
 [18] copyto!
    @ ~/.julia/packages/GPUArrays/ymNij/src/host/broadcast.jl:63 [inlined]
 [19] copyto!
    @ ./broadcast.jl:936 [inlined]
 [20] copy
    @ ~/.julia/packages/GPUArrays/ymNij/src/host/broadcast.jl:47 [inlined]
 [21] materialize(bc::Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{1}, Nothing, typeof(div), Tuple{CuArray{Int128, 1}, CuArray{Int128, 1}}})
    @ Base.Broadcast ./broadcast.jl:883
 [22] top-level scope
    @ REPL[12]:1
 [23] top-level scope
    @ ~/.julia/packages/CUDA/mVgLI/src/initialization.jl:52

Is this error related to the architecture of my GPU or just not binded correctly?
My GPU is Nvidia V100, CUDA version is 3.3, julia version is 1.6.0.

2

Int128 isn’t natively supported, and LLVM relies on intrinsics for many operations. We don’t currently have these intrinsics (implemented or linked in from another library), resulting in the error you encountered.

1 Like
3

I see, thanks for explaining. I will switch to alternative implementations.

4

You can try providing the intrinsic: a C implementation can be found in compiler-rt, compiler-rt/divti3.c at master · llvm-mirror/compiler-rt · GitHub, and to add it to the runtime you’d put that implementation in GPUCompiler.jl, GPUCompiler.jl/runtime.jl at a2ca4b2cb36206fb43da6dd71caed16dd8b9d49a · JuliaGPU/GPUCompiler.jl · GitHub.

2 Likes