Unexpected error: ccall requires compiler when using QR

[mfherbst]

I wanted to use the qr routines added in #456 and here is what I get:

using AMDGPU
using LinearAlgebra

A = ROCArray(rand(ComplexF64, 8, 4))
qr(A)

Stacktrace:

Internal error: encountered unexpected error during compilation of geqrf!:
TypeError(func=:ccall, context="", expected=Symbol, got=nothing)

[... long stacktrace ...]

unknown function (ip: 0x401098)
ERROR: `ccall` requires the compiler
Stacktrace:
  [1] rocblas_create_handle(handle::Base.RefValue{Ptr{Nothing}})
    @ AMDGPU.rocBLAS ~/.julia/packages/AMDGPU/1A9UJ/src/blas/librocblas.jl:188
  [2] macro expansion
    @ ~/.julia/packages/AMDGPU/1A9UJ/src/utils.jl:182 [inlined]
  [3] create_handle()
    @ AMDGPU.rocBLAS ~/.julia/packages/AMDGPU/1A9UJ/src/blas/rocBLAS.jl:28
  [4] (::AMDGPU.var"#28#35")()
    @ AMDGPU ~/.julia/packages/AMDGPU/1A9UJ/src/cache.jl:103
  [5] (::AMDGPU.var"#19#22")()
    @ AMDGPU ~/.julia/packages/AMDGPU/1A9UJ/src/cache.jl:29
  [6] lock(f::AMDGPU.var"#19#22"{AMDGPU.var"#28#35"{typeof(AMDGPU.rocBLAS.create_handle)}, HandleCache{HIPContext, Ptr{Nothing}}, HIPContext}, l::ReentrantLock)
    @ Base ./lock.jl:229
  [7] check_cache
    @ ~/.julia/packages/AMDGPU/1A9UJ/src/cache.jl:27 [inlined]
  [8] pop!
    @ ~/.julia/packages/AMDGPU/1A9UJ/src/cache.jl:48 [inlined]
  [9] (::AMDGPU.var"#new_state#34"{Ptr{Nothing}, HandleCache{HIPContext, Ptr{Nothing}}, typeof(AMDGPU.rocBLAS.create_handle), AMDGPU.rocBLAS.var"#15#16"})(tls::AMDGPU.TaskLocalState)
    @ AMDGPU ~/.julia/packages/AMDGPU/1A9UJ/src/cache.jl:102
 [10] #32
    @ ~/.julia/packages/AMDGPU/1A9UJ/src/cache.jl:115 [inlined]
 [11] get!(default::AMDGPU.var"#32#39"{AMDGPU.var"#new_state#34"{Ptr{Nothing}, HandleCache{HIPContext, Ptr{Nothing}}, typeof(AMDGPU.rocBLAS.create_handle), AMDGPU.rocBLAS.var"#15#16"}, AMDGPU.TaskLocalState}, h::Dict{HIPContext, NamedTuple{(:handle, :stream), Tuple{Ptr{Nothing}, HIPStream}}}, key::HIPContext)
    @ Base ./dict.jl:468
 [12] library_state(library_key::Symbol, #unused#::Type{Ptr{Nothing}}, idle_handles::HandleCache{HIPContext, Ptr{Nothing}}, create_handle::typeof(AMDGPU.rocBLAS.create_handle), destroy_handle::Function, set_stream::AMDGPU.rocBLAS.var"#15#16")
    @ AMDGPU ~/.julia/packages/AMDGPU/1A9UJ/src/cache.jl:115
 [13] lib_state
    @ ~/.julia/packages/AMDGPU/1A9UJ/src/blas/rocBLAS.jl:39 [inlined]
 [14] handle()
    @ AMDGPU.rocBLAS ~/.julia/packages/AMDGPU/1A9UJ/src/blas/rocBLAS.jl:44
 [15] geqrf!(A::ROCMatrix{ComplexF64})
    @ AMDGPU.rocSOLVER ~/.julia/packages/AMDGPU/1A9UJ/src/solver/highlevel.jl:13
 [16] qr!
    @ ~/.julia/packages/AMDGPU/1A9UJ/src/solver/highlevel.jl:206 [inlined]
 [17] #qr#85
    @ ~/julia-1.9.2/share/julia/stdlib/v1.9/LinearAlgebra/src/qr.jl:428 [inlined]
 [18] qr(::ROCMatrix{ComplexF64, AMDGPU.Runtime.Mem.HIPBuffer})
    @ LinearAlgebra ~/julia-1.9.2/share/julia/stdlib/v1.9/LinearAlgebra/src/qr.jl:425
 [19] top-level scope
    @ REPL[4]:1

[pxl-th]

That's weird, can you post AMDGPU.versioninfo() & versioninfo()?
Maybe that's some configuration issue.

Above example works on my machine:

julia> using LinearAlgebra, AMDGPU
[ Info: Navi 2 GPU detected, setting ENV variable: `HSA_OVERRIDE_GFX_VERSION=10.3.0`.

julia> A = ROCArray(rand(ComplexF64, 8, 4));

julia> qr(A)
QR{ComplexF64, ROCMatrix{ComplexF64, AMDGPU.Runtime.Mem.HIPBuffer}, ROCVector{ComplexF64, AMDGPU.Runtime.Mem.HIPBuffer}}
Q factor:
8×8 LinearAlgebra.QRPackedQ{ComplexF64, ROCMatrix{ComplexF64, AMDGPU.Runtime.Mem.HIPBuffer}, ROCVector{ComplexF64, AMDGPU.Runtime.Mem.HIPBuffer}}:
  -0.263534-0.383082im    0.201817+0.300202im   -0.0609926+0.323185im    …   -0.391813-0.115354im   -0.0163812+0.0353206im
  -0.149431-0.135789im  -0.0804106-0.393071im    -0.192847-0.104253im         -0.30923+0.306449im     -0.49432-0.225866im
  -0.361372-0.039774im   -0.050347-0.431445im    -0.320076+0.358908im         0.160851-0.228831im      0.30265+0.266208im
 -0.0163679-0.211219im   -0.372776+0.0878281im    0.459565-0.0414137im        0.118538-0.11094im      -0.19236+0.122445im
  -0.322477-0.365644im     0.12508+0.24006im      0.209779+0.270177im        0.0764277-0.0450508im   -0.118406+0.0710881im
  -0.130716-0.239194im   -0.279746-0.21806im      0.159488+0.00837835im  …   -0.214403-0.0708149im   -0.106042-0.185207im
  -0.167977-0.261044im    0.121521-0.15146im      0.266802-0.165913im         0.635625-0.0514819im  -0.0298735-0.3116im
   -0.20606-0.339365im   -0.360917-0.0759747im   -0.114503-0.381134im       -0.0908846+0.255466im     0.570497-0.0241447im
R factor:
4×4 ROCMatrix{ComplexF64, AMDGPU.Runtime.Mem.HIPBuffer}:
 -2.35654+0.0im  -1.79344+0.0237188im  -1.42258+0.412774im    -1.58647-0.0509862im
      0.0+0.0im  -1.66205+0.0im        -1.00484+0.0649698im  -0.696264-0.023673im
      0.0+0.0im       0.0+0.0im        0.923757+0.0im         0.460948+0.441239im
      0.0+0.0im       0.0+0.0im             0.0+0.0im         0.915658+0.0im

[mfherbst]

Here you go:

julia> AMDGPU.versioninfo()
ROCm provided by: system
[+] HSA Runtime v1.1.0
    @ /opt/rocm/hsa/lib/libhsa-runtime64.so
[+] ld.lld
    @ /opt/rocm/llvm/bin/ld.lld
[+] ROCm-Device-Libs
    @ /opt/rocm/amdgcn/bitcode
[+] HIP Runtime v5.504.22804
    @ /opt/rocm/hip/lib/libamdhip64.so
[-] rocBLAS 
[-] rocSOLVER 
[-] rocALUTION
[-] rocSPARSE
[-] rocRAND 
[+] rocFFT v1.0.21
    @ /opt/rocm/lib/librocfft.so
[-] MIOpen 

HIP Devices [2]
    1. HIPDevice(name="AMD Radeon RX Vega", id=1)
    2. HIPDevice(name="AMD Radeon RX 480 Graphics", id=2)

julia> versioninfo()
Julia Version 1.9.2
Commit e4ee485e909 (2023-07-05 09:39 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 4 × Intel(R) Core(TM) i5-4590 CPU @ 3.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, haswell)
  Threads: 1 on 4 virtual cores

[mfherbst]

Thanks for looking into it!

[mfherbst]

Ah it seems rocsolver is not installed ... that's weird. But maybe there could be a more graceful error message?

[pxl-th]

But maybe there could be a more graceful error message?

Agree, will add it.