Compilation issues with `Flux.softmax()` and Julia v1.12.0-beta2

[nstiurca]

I tried running Flux.jl with AMDGPU backend under Julia v1.12.0-beta2, but pretty quickly I ran into issues with LLVM errors when compiling kernels. I already filed JuliaLang/julia#58310 but they suggested I come here instead. Copying the rest of that issue here for simplicity.

Here's a minimal example to reproduce, which notably works on Julia v1.11:

nicu@blackstash:~$ julia +1.12 --startup-file=no
               _
   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.12.0-beta2 (2025-04-25)
 _/ |\__'_|_|_|\__'_|  |  Official https://julialang.org release
|__/                   |

julia> using AMDGPU; using Flux: softmax

julia> data = ROCArray([0.5])
1-element ROCArray{Float64, 1, AMDGPU.Runtime.Mem.HIPBuffer}:
 0.5

julia> softmax(data)
ERROR: LLVM error: Cannot select: 0x232d3d20: f64 = fmaximum nnan ninf nsz arcp contract afn reassoc # D:1 0x23398dd0, 0x226fed50, fastmath.jl:171 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ]
  0x23398dd0: f64,ch = CopyFromReg 0x2325e260, Register:f64 %23, fastmath.jl:171 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ]
    0x226fece0: f64 = Register %23
  0x226fed50: f64 = bitcast # D:1 0x241c46c0, /home/nicu/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ none:0 @[ none:0 @[ /home/nicu/.julia/packages/LLVM/xTJfF/src/interop/pointer.jl:85 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/device/gcn/array.jl:81 @[ abstractarray.jl:1366 @[ abstractarray.jl:1342 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ]
    0x241c46c0: v2i32,ch = load<(load (s64) from %ir.gep.peel, !tbaa !199, addrspace 1)> # D:1 0x2325e260, 0x232d4180, undef:i64, /home/nicu/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ none:0 @[ none:0 @[ /home/nicu/.julia/packages/LLVM/xTJfF/src/interop/pointer.jl:85 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/device/gcn/array.jl:81 @[ abstractarray.jl:1366 @[ abstractarray.jl:1342 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ]
      0x232d4180: i64 = add # D:1 0x23398f90, 0x22a655e0, /home/nicu/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ none:0 @[ none:0 @[ /home/nicu/.julia/packages/LLVM/xTJfF/src/interop/pointer.jl:85 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/device/gcn/array.jl:81 @[ abstractarray.jl:1366 @[ abstractarray.jl:1342 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ]
        0x23398f90: i64 = add 0x241c4a40, Constant:i64<-8>, int.jl:87 @[ int.jl:863 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:65 ] ]
          0x241c4a40: i64 = bitcast 0x22a65340
            0x22a65340: v2i32,ch = load<(dereferenceable invariant load (s64) from %ir."As[1]::ROCDeviceArray.kernarg.offset" + 8, basealign 16, addrspace 4)> 0x2325e260, 0x22a65260, undef:i64
              0x22a65260: i64 = add 0x23398eb0, Constant:i64<152>
                0x23398eb0: i64,ch = CopyFromReg 0x2325e260, Register:i64 %0
                  0x22a65810: i64 = Register %0
                0x23e944c0: i64 = Constant<152>
              0x241c42d0: i64 = undef
          0x23399000: i64 = Constant<-8>
        0x22a655e0: i64 = shl # D:1 0x226fd870, Constant:i32<3>, /home/nicu/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ none:0 @[ none:0 @[ /home/nicu/.julia/packages/LLVM/xTJfF/src/interop/pointer.jl:85 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/device/gcn/array.jl:81 @[ abstractarray.jl:1366 @[ abstractarray.jl:1342 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ]
          0x226fd870: i64 = bitcast # D:1 0x226fdcd0, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
            0x226fdcd0: v2i32 = BUILD_VECTOR # D:1 0x226fdf00, 0x226fdd40, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
              0x226fdf00: i32 = select # D:1 0x232d3fc0, 0x226fe5e0, 0x226fdfc0, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
                0x232d3fc0: i1 = setcc # D:1 0x23c5c8c0, 0x23398c10, setgt:ch, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
                  0x23c5c8c0: i64,ch = CopyFromReg # D:1 0x2325e260, Register:i64 %28, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]

                  0x23398c10: i64,ch = CopyFromReg 0x2325e260, Register:i64 %15, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]

                0x226fe5e0: i32 = extract_vector_elt # D:1 0x23e95090, Constant:i32<0>, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
                  0x23e95090: v2i32 = bitcast # D:1 0x23c5c8c0, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]

                  0x232d4c70: i32 = Constant<0>
                0x226fdfc0: i32 = extract_vector_elt 0x241c4490, Constant:i32<0>, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
                  0x241c4490: v2i32 = bitcast 0x23398c10, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]

                  0x232d4c70: i32 = Constant<0>
              0x226fdd40: i32 = select # D:1 0x232d3fc0, 0x226fe880, 0x226fedc0, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
                0x232d3fc0: i1 = setcc # D:1 0x23c5c8c0, 0x23398c10, setgt:ch, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
                  0x23c5c8c0: i64,ch = CopyFromReg # D:1 0x2325e260, Register:i64 %28, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]

                  0x23398c10: i64,ch = CopyFromReg 0x2325e260, Register:i64 %15, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]

                0x226fe880: i32 = extract_vector_elt # D:1 0x23e95090, Constant:i32<1>, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
                  0x23e95090: v2i32 = bitcast # D:1 0x23c5c8c0, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]

                  0x22a653b0: i32 = Constant<1>
                0x226fedc0: i32 = extract_vector_elt 0x241c4490, Constant:i32<1>, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
                  0x241c4490: v2i32 = bitcast 0x23398c10, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:385 @[ multidimensional.jl:129 @[ /home/nicu/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]

                  0x22a653b0: i32 = Constant<1>
          0x226fd090: i32 = Constant<3>
      0x241c42d0: i64 = undef
In function: _Z24partial_mapreduce_device8identity8max_fast7Float6416CartesianIndicesILi1E5TupleI5OneToI5Int64EEES8_14ROCDeviceArrayIS1_Li2ELi1EES9_IS1_Li1ELi1EE
Stacktrace:
  [1] handle_error(reason::Cstring)
    @ LLVM ~/.julia/packages/LLVM/xTJfF/src/core/context.jl:194
  [2] LLVMTargetMachineEmitToMemoryBuffer
    @ ~/.julia/packages/LLVM/xTJfF/lib/18/libLLVM.jl:11531 [inlined]
  [3] emit(tm::LLVM.TargetMachine, mod::LLVM.Module, filetype::LLVM.API.LLVMCodeGenFileType)
    @ LLVM ~/.julia/packages/LLVM/xTJfF/src/targetmachine.jl:118
  [4] mcgen(job::GPUCompiler.CompilerJob, mod::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/mcgen.jl:75
  [5] macro expansion
    @ ~/.julia/packages/Tracy/GcShf/src/tracepoint.jl:158 [inlined]
  [6] macro expansion
    @ ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:404 [inlined]
  [7] macro expansion
    @ ~/.julia/packages/Tracy/GcShf/src/tracepoint.jl:158 [inlined]
  [8] emit_asm(job::GPUCompiler.CompilerJob, ir::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:401
  [9] compile_unhooked(output::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:115
 [10] compile_unhooked
    @ ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:80 [inlined]
 [11] compile(target::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:67
 [12] compile
    @ ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:55 [inlined]
 [13] #hipcompile##0
    @ ~/.julia/packages/AMDGPU/6s4nD/src/compiler/codegen.jl:194 [inlined]
 [14] JuliaContext(f::AMDGPU.Compiler.var"#hipcompile##0#hipcompile##1"{GPUCompiler.CompilerJob{…}}; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:34
 [15] JuliaContext(f::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:25
 [16] hipcompile(job::GPUCompiler.CompilerJob)
    @ AMDGPU.Compiler ~/.julia/packages/AMDGPU/6s4nD/src/compiler/codegen.jl:193
 [17] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(AMDGPU.Compiler.hipcompile), linker::typeof(AMDGPU.Compiler.hiplink))
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/execution.jl:245
 [18] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/execution.jl:159
 [19] macro expansion
    @ ~/.julia/packages/AMDGPU/6s4nD/src/compiler/codegen.jl:161 [inlined]
 [20] macro expansion
    @ ./lock.jl:376 [inlined]
 [21] hipfunction(f::typeof(AMDGPU.partial_mapreduce_device), tt::Type{Tuple{…}}; kwargs::@Kwargs{})
    @ AMDGPU.Compiler ~/.julia/packages/AMDGPU/6s4nD/src/compiler/codegen.jl:155
 [22] hipfunction
    @ ~/.julia/packages/AMDGPU/6s4nD/src/compiler/codegen.jl:154 [inlined]
 [23] macro expansion
    @ ~/.julia/packages/AMDGPU/6s4nD/src/highlevel.jl:153 [inlined]
 [24] mapreducedim!(f::typeof(identity), op::typeof(Base.FastMath.max_fast), R::ROCArray{…}, A::ROCArray{…}; init::Float64)
    @ AMDGPU ~/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:134
 [25] mapreducedim!
    @ ~/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:86 [inlined]
 [26] _mapreduce(f::typeof(identity), op::typeof(Base.FastMath.max_fast), As::ROCArray{…}; dims::Int64, init::Float64)
    @ GPUArrays ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:76
 [27] _mapreduce
    @ ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:33 [inlined]
 [28] mapreduce
    @ ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:28 [inlined]
 [29] reduce
    @ ./reducedim.jl:375 [inlined]
 [30] fast_maximum
    @ ~/.julia/packages/NNlib/CGMj3/src/softmax.jl:92 [inlined]
 [31] softmax!(out::ROCArray{Float64, 1, AMDGPU.Runtime.Mem.HIPBuffer}, x::ROCArray{Float64, 1, AMDGPU.Runtime.Mem.HIPBuffer}; dims::Int64)
    @ NNlib ~/.julia/packages/NNlib/CGMj3/src/softmax.jl:61
 [32] softmax!
    @ ~/.julia/packages/NNlib/CGMj3/src/softmax.jl:60 [inlined]
 [33] softmax(x::ROCArray{Float64, 1, AMDGPU.Runtime.Mem.HIPBuffer}; dims::Int64)
    @ NNlib ~/.julia/packages/NNlib/CGMj3/src/softmax.jl:56
 [34] softmax(x::ROCArray{Float64, 1, AMDGPU.Runtime.Mem.HIPBuffer})
    @ NNlib ~/.julia/packages/NNlib/CGMj3/src/softmax.jl:56
 [35] top-level scope
    @ REPL[3]:1
Some type information was truncated. Use `show(err)` to see complete types.

Expected behavior:

nicu@blackstash:~$ julia +1.11 --startup-file=no
               _
   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.11.5 (2025-04-14)
 _/ |\__'_|_|_|\__'_|  |  Official https://julialang.org/ release
|__/                   |

julia> using AMDGPU; using Flux: softmax

julia> data = ROCArray([0.5])
1-element ROCArray{Float64, 1, AMDGPU.Runtime.Mem.HIPBuffer}:
 0.5

julia> softmax(data)
1-element ROCArray{Float64, 1, AMDGPU.Runtime.Mem.HIPBuffer}:
 1.0

For reference,

julia> versioninfo()
Julia Version 1.12.0-beta2
Commit dd74040f22b (2025-04-25 12:03 UTC)
Build Info:
  Official https://julialang.org release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 16 × AMD Ryzen 9 7940HS w/ Radeon 780M Graphics
  WORD_SIZE: 64
  LLVM: libLLVM-18.1.7 (ORCJIT, znver4)
  GC: Built with stock GC
Threads: 16 default, 1 interactive, 16 GC (on 16 virtual cores)
Environment:
  JULIA_PROJECT = @.
  JULIA_NUM_THREADS = auto
  JULIA_PKG_DEVDIR = /home/nicu/src

julia> AMDGPU.device()
┌────┬─────────────────────┬──────────┬───────────┬────────────┬────────────────
│ Id │                Name │ GCN arch │ Wavefront │     Memory │ Shared Memory ⋯
├────┼─────────────────────┼──────────┼───────────┼────────────┼────────────────
│  1 │ AMD Radeon Graphics │  gfx1100 │        32 │ 27.364 GiB │    64.000 KiB ⋯
└────┴─────────────────────┴──────────┴───────────┴────────────┴────────────────

It's also worth noting that my iGPU is not officially supported in ROCm v6.4, so I used a well-known-hack to even get it working on Julia 1.11. The 780m in my APU is actually a gfx1103, but I have to convince ROCm to treat it as a gfx1100 with

export HSA_OVERRIDE_GFX_VERSION=11.0.0

So there it is. I understand I'm using an unsupported hardware/software combination, but this did work in Julia v1.11 so I'm hoping this can continue to work in v1.12 as well. FWIW, the entire Flux.jl quickstart works well for me if I just replace the using CUDA; device = gpu_device() bit with using AMDGPU; device = roc.

[pxl-th]

Hi. Thanks for reporting that.
I'm away from my AMD GPU machine for the next few day, so can you confirm for me few things:

1. make sure you are on GPUCompiler@1.4.0 & AMDGPU@1.3.0.
2. kernel that causes error is this one, can you try executing following version and see if it also errors:

julia> x = AMDGPU.rand(Float64, 128, 2);

julia> fast_maximum(x::AbstractArray{T}; dims) where {T} = reduce(max, x; dims, init = float(T)(-Inf))

julia> fast_maximum(x; dims=1)

[luraess]

Hi. Thanks for reporting that. I'm away from my AMD GPU machine for the next few day, so can you confirm for me few things:

1. make sure you are on GPUCompiler@1.4.0 & AMDGPU@1.3.0.
2. kernel that causes error is this one, can you try executing following version and see if it also errors:

julia> x = AMDGPU.rand(Float64, 128, 2);

julia> fast_maximum(x::AbstractArray{T}; dims) where {T} = reduce(max, x; dims, init = float(T)(-Inf))

julia> fast_maximum(x; dims=1)

Testing this on the CI machine (ROCm 6.4, Ubuntu 24.04, AMDGPU 1.3.0, GPUCompiler 1.4.0), it works. However, adding the @fastmath macro as in FluxML implementation reproduces the error from the issue.

[mcabbott]

Does it also fail with fast_maximum(x::AbstractArray; dims) = @fastmath maximum(x; dims)? On an Array that's the same thing, but some chance GPU stuff overloads Base.FastMath.maximum_fast.

[luraess]

fast_maximum(x::AbstractArray; dims) = @fastmath maximum(x; dims)

See

julia> x = AMDGPU.rand(Float64, 128, 2);

julia> fast_maximum(x::AbstractArray; dims) = @fastmath maximum(x; dims)
fast_maximum (generic function with 1 method)

julia> fast_maximum(x; dims=1)
ERROR: GPUArrays.jl needs to know the neutral element for your operator `max_fast`.
Please pass it as an explicit argument to `GPUArrays.mapreducedim!`,
or register it globally by defining `GPUArrays.neutral_element(::typeof(max_fast), T)`.
Stacktrace:
 [1] error(s::String)
   @ Base ./error.jl:44
 [2] neutral_element(op::Function, T::Type)
   @ GPUArrays ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:13
 [3] _mapreduce(f::typeof(identity), op::typeof(Base.FastMath.max_fast), As::ROCArray{…}; dims::Int64, init::Nothing)
   @ GPUArrays ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:42
 [4] mapreduce(::Function, ::Function, ::ROCArray{Float64, 2, AMDGPU.Runtime.Mem.HIPBuffer}; dims::Int64, init::Nothing)
   @ GPUArrays ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:28
 [5] mapreduce
   @ ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:28 [inlined]
 [6] #reduce#729
   @ ./reducedim.jl:375 [inlined]
 [7] maximum_fast(a::ROCArray{Float64, 2, AMDGPU.Runtime.Mem.HIPBuffer}; kw::@Kwargs{dims::Int64})
   @ Base.FastMath ./fastmath.jl:398
 [8] fast_maximum(x::ROCArray{Float64, 2, AMDGPU.Runtime.Mem.HIPBuffer}; dims::Int64)
   @ Main ./REPL[3]:1
 [9] top-level scope
   @ REPL[4]:1
Some type information was truncated. Use `show(err)` to see complete types.

julia> maximum(x; dims=1)
1×2 ROCArray{Float64, 2, AMDGPU.Runtime.Mem.HIPBuffer}:
 0.995202  0.973682

julia> @fastmath maximum(x; dims=1)
ERROR: GPUArrays.jl needs to know the neutral element for your operator `max_fast`.
Please pass it as an explicit argument to `GPUArrays.mapreducedim!`,
or register it globally by defining `GPUArrays.neutral_element(::typeof(max_fast), T)`.
Stacktrace:
 [1] error(s::String)
   @ Base ./error.jl:44
 [2] neutral_element(op::Function, T::Type)
   @ GPUArrays ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:13
 [3] _mapreduce(f::typeof(identity), op::typeof(Base.FastMath.max_fast), As::ROCArray{…}; dims::Int64, init::Nothing)
   @ GPUArrays ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:42
 [4] mapreduce(::Function, ::Function, ::ROCArray{Float64, 2, AMDGPU.Runtime.Mem.HIPBuffer}; dims::Int64, init::Nothing)
   @ GPUArrays ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:28
 [5] mapreduce
   @ ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:28 [inlined]
 [6] #reduce#729
   @ ./reducedim.jl:375 [inlined]
 [7] maximum_fast(a::ROCArray{Float64, 2, AMDGPU.Runtime.Mem.HIPBuffer}; kw::@Kwargs{dims::Int64})
   @ Base.FastMath ./fastmath.jl:398
 [8] top-level scope
   @ REPL[6]:1
Some type information was truncated. Use `show(err)` to see complete types.

[mcabbott]

Does defining GPUArrays.neutral_element(::typeof(Base.FastMath.max_fast), ::Type{T}) where {T} = Base.typemin(T) solve that, or produce the earlier error?

[Edit: fully qualified the names]

[luraess]

max_fast operator seems to be unknown at this level

[nstiurca]

Thanks for the quick investigation. I can confirm all the same results as @luraess on my machine.

[luraess]

Does defining GPUArrays.neutral_element(::typeof(Base.FastMath.max_fast), ::Type{T}) where {T} = Base.typemin(T) solve that, or produce the earlier error?

[Edit: fully qualified the names]

Thanks. Adding this changes the error making it more similar to the earlier one

julia> using AMDGPU, GPUArrays

julia> fast_maximum(x::AbstractArray; dims) = @fastmath maximum(x; dims)
fast_maximum (generic function with 1 method)

julia> GPUArrays.neutral_element(::typeof(Base.FastMath.max_fast), ::Type{T}) where {T} = Base.typemin(T)

julia> x = AMDGPU.rand(Float64, 128, 2);

julia> fast_maximum(x; dims=1)
ERROR: LLVM error: Cannot select: 0x32ee1db0: f64 = fmaximum nnan ninf nsz arcp contract afn reassoc # D:1 0x32847700, 0x31472da0, fastmath.jl:171 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ]
  0x32847700: f64,ch = CopyFromReg 0x3020b570, Register:f64 %65, fastmath.jl:171 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ]
    0x32fbdb90: f64 = Register %65
  0x31472da0: f64 = bitcast # D:1 0x32be7ed0, /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ none:0 @[ none:0 @[ /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/pointer.jl:85 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/device/gcn/array.jl:81 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ]
    0x32be7ed0: v2i32,ch = load<(load (s64) from %ir.205, !tbaa !308, addrspace 1)> # D:1 0x32ee15d0, 0x32fbe4c0, undef:i64, /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ none:0 @[ none:0 @[ /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/pointer.jl:85 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/device/gcn/array.jl:81 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ]
      0x32fbe4c0: i64 = add # D:1 0x31472940, Constant:i64<-8>, /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ none:0 @[ none:0 @[ /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/pointer.jl:85 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/device/gcn/array.jl:81 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ]
        0x31472940: i64 = add # D:1 0x328d71d0, 0x32be81e0, /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ none:0 @[ none:0 @[ /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/pointer.jl:85 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/device/gcn/array.jl:81 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ]
          0x328d71d0: i64 = add # D:1 0x3020e870, 0x31473350, /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ none:0 @[ none:0 @[ /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/pointer.jl:85 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/device/gcn/array.jl:81 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ]
            0x3020e870: i64 = bitcast 0x3020eb80
              0x3020eb80: v2i32,ch = load<(dereferenceable invariant load (s64) from %ir."As[1]::ROCDeviceArray.kernarg.offset" + 16, addrspace 4)> 0x3020b570, 0x32be7f40, undef:i64
                0x32be7f40: i64 = add 0x32ee16b0, Constant:i64<184>
                  0x32ee16b0: i64,ch = CopyFromReg 0x3020b570, Register:i64 %0

                  0x32ee1cd0: i64 = Constant<184>
                0x32be8a30: i64 = undef
            0x31473350: i64 = shl # D:1 0x32847c40, Constant:i32<3>, /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ none:0 @[ none:0 @[ /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/pointer.jl:85 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/device/gcn/array.jl:81 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ]
              0x32847c40: i64 = bitcast # D:1 0x31472be0, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:386 @[ multidimensional.jl:129 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
                0x31472be0: v2i32 = BUILD_VECTOR # D:1 0x32ee1e90, 0x32be8800, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:386 @[ multidimensional.jl:129 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]
                  0x32ee1e90: i32 = select # D:1 0x328d6de0, 0x32fbe530, 0x3020e4f0, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:386 @[ multidimensional.jl:129 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]



                  0x32be8800: i32 = select # D:1 0x328d6de0, 0x2c439630, 0x328d6ad0, essentials.jl:799 @[ promotion.jl:643 @[ tuple.jl:386 @[ multidimensional.jl:129 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:68 ] ] ] ]



              0x3020ebf0: i32 = Constant<3>
          0x32be81e0: i64 = shl # D:1 0x328d6830, Constant:i32<3>, /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ none:0 @[ none:0 @[ /home/lraess/.julia/packages/LLVM/xTJfF/src/interop/pointer.jl:85 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/device/gcn/array.jl:81 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ]
            0x328d6830: i64 = bitcast # D:1 0x314734a0, int.jl:88 @[ abstractarray.jl:3070 @[ abstractarray.jl:3070 @[ abstractarray.jl:3054 @[ abstractarray.jl:3038 @[ abstractarray.jl:1377 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ] ]
              0x314734a0: v2i32 = BUILD_VECTOR # D:1 0x2c439550, 0x32ee1c60, int.jl:88 @[ abstractarray.jl:3070 @[ abstractarray.jl:3070 @[ abstractarray.jl:3054 @[ abstractarray.jl:3038 @[ abstractarray.jl:1377 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ] ]
                0x2c439550: i32,i32 = umul_lohi # D:1 0x31473740, 0x32ee12c0, int.jl:88 @[ abstractarray.jl:3070 @[ abstractarray.jl:3070 @[ abstractarray.jl:3054 @[ abstractarray.jl:3038 @[ abstractarray.jl:1377 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ] ]
                  0x31473740: i32 = truncate # D:1 0x328475b0, int.jl:88 @[ abstractarray.jl:3070 @[ abstractarray.jl:3070 @[ abstractarray.jl:3054 @[ abstractarray.jl:3038 @[ abstractarray.jl:1377 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ] ]

                  0x32ee12c0: i32 = extract_vector_elt 0x32ee1640, Constant:i32<0>, int.jl:88 @[ abstractarray.jl:3070 @[ abstractarray.jl:3070 @[ abstractarray.jl:3054 @[ abstractarray.jl:3038 @[ abstractarray.jl:1377 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ] ]


                0x32ee1c60: i32 = add # D:1 0x32847620, 0x32ee1170, int.jl:88 @[ abstractarray.jl:3070 @[ abstractarray.jl:3070 @[ abstractarray.jl:3054 @[ abstractarray.jl:3038 @[ abstractarray.jl:1377 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ] ]
                  0x32847620: i32 = add # D:1 0x2c439550:1, 0x328480a0, int.jl:88 @[ abstractarray.jl:3070 @[ abstractarray.jl:3070 @[ abstractarray.jl:3054 @[ abstractarray.jl:3038 @[ abstractarray.jl:1377 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ] ]


                  0x32ee1170: i32 = mul # D:1 0x2c4392b0, 0x32ee12c0, int.jl:88 @[ abstractarray.jl:3070 @[ abstractarray.jl:3070 @[ abstractarray.jl:3054 @[ abstractarray.jl:3038 @[ abstractarray.jl:1377 @[ abstractarray.jl:1371 @[ abstractarray.jl:1342 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:37 @[ /home/lraess/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:69 ] ] ] ] ] ] ] ] ]


            0x3020ebf0: i32 = Constant<3>
        0x31472d30: i64 = Constant<-8>
      0x32be8a30: i64 = undef
In function: _Z24partial_mapreduce_device8identity8max_fast7Float6416CartesianIndicesILi2E5TupleI5OneToI5Int64ES6_EES8_14ROCDeviceArrayIS1_Li3ELi1EES9_IS1_Li2ELi1EE
Stacktrace:
  [1] handle_error(reason::Cstring)
    @ LLVM ~/.julia/packages/LLVM/xTJfF/src/core/context.jl:194
  [2] LLVMTargetMachineEmitToMemoryBuffer
    @ ~/.julia/packages/LLVM/xTJfF/lib/18/libLLVM.jl:11531 [inlined]
  [3] emit(tm::LLVM.TargetMachine, mod::LLVM.Module, filetype::LLVM.API.LLVMCodeGenFileType)
    @ LLVM ~/.julia/packages/LLVM/xTJfF/src/targetmachine.jl:118
  [4] mcgen(job::GPUCompiler.CompilerJob, mod::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/mcgen.jl:75
  [5] macro expansion
    @ ~/.julia/packages/Tracy/GcShf/src/tracepoint.jl:158 [inlined]
  [6] macro expansion
    @ ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:404 [inlined]
  [7] macro expansion
    @ ~/.julia/packages/Tracy/GcShf/src/tracepoint.jl:158 [inlined]
  [8] emit_asm(job::GPUCompiler.CompilerJob, ir::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:401
  [9] compile_unhooked(output::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:115
 [10] compile_unhooked
    @ ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:80 [inlined]
 [11] compile(target::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:67
 [12] compile
    @ ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:55 [inlined]
 [13] #hipcompile##0
    @ ~/.julia/packages/AMDGPU/6s4nD/src/compiler/codegen.jl:194 [inlined]
 [14] JuliaContext(f::AMDGPU.Compiler.var"#hipcompile##0#hipcompile##1"{GPUCompiler.CompilerJob{…}}; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:34
 [15] JuliaContext(f::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/driver.jl:25
 [16] hipcompile(job::GPUCompiler.CompilerJob)
    @ AMDGPU.Compiler ~/.julia/packages/AMDGPU/6s4nD/src/compiler/codegen.jl:193
 [17] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(AMDGPU.Compiler.hipcompile), linker::typeof(AMDGPU.Compiler.hiplink))
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/execution.jl:245
 [18] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/1cGqD/src/execution.jl:159
 [19] macro expansion
    @ ~/.julia/packages/AMDGPU/6s4nD/src/compiler/codegen.jl:161 [inlined]
 [20] macro expansion
    @ ./lock.jl:376 [inlined]
 [21] hipfunction(f::typeof(AMDGPU.partial_mapreduce_device), tt::Type{Tuple{…}}; kwargs::@Kwargs{})
    @ AMDGPU.Compiler ~/.julia/packages/AMDGPU/6s4nD/src/compiler/codegen.jl:155
 [22] hipfunction
    @ ~/.julia/packages/AMDGPU/6s4nD/src/compiler/codegen.jl:154 [inlined]
 [23] macro expansion
    @ ~/.julia/packages/AMDGPU/6s4nD/src/highlevel.jl:153 [inlined]
 [24] mapreducedim!(f::typeof(identity), op::typeof(Base.FastMath.max_fast), R::ROCArray{…}, A::ROCArray{…}; init::Float64)
    @ AMDGPU ~/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:134
 [25] mapreducedim!
    @ ~/.julia/packages/AMDGPU/6s4nD/src/kernels/mapreduce.jl:86 [inlined]
 [26] _mapreduce(f::typeof(identity), op::typeof(Base.FastMath.max_fast), As::ROCArray{…}; dims::Int64, init::Nothing)
    @ GPUArrays ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:76
 [27] _mapreduce
    @ ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:33 [inlined]
 [28] mapreduce
    @ ~/.julia/packages/GPUArrays/uiVyU/src/host/mapreduce.jl:28 [inlined]
 [29] reduce
    @ ./reducedim.jl:375 [inlined]
 [30] maximum_fast
    @ ./fastmath.jl:398 [inlined]
 [31] #fast_maximum#1
    @ ./REPL[2]:1 [inlined]
 [32] top-level scope
    @ REPL[6]:1
Some type information was truncated. Use `show(err)` to see complete types.

[giordano]

@pxl-th if it helps, this is the output of @device_code dir="./devcode" fast_maximum(x; dims=1): devcode-maximum.tar.gz

[pxl-th]

TL;DR: Julia 1.12 moved to llvm.minimum/maximum for min/max, but AMDGPU supports it since LLVM 19.

Ref: JuliaLang/julia#56371

MWE:

using AMDGPU

function ker!(x)
    @inbounds x[1] = @fastmath max(x[1], 0.0)
    return
end

function main()
    x = ROCArray(ones(Float64, 1))
    @roc ker!(x)
    return
end
main()