search

JuliaGPU / KernelAbstractions.jl

Heterogeneous programming in Julia
MIT License
363 stars 65 forks source link

amd gpu give different results when nested loop is used #517

Open ww1g11 opened 1 week ago

ww1g11 commented 1 week ago

Hi, I noticed that the following script produces different results depending on the backend. On my machine, the output is:

cpu: [18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0;;;]
cuda: [18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0;;;]
amd: [6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0;;;]

Is there a mistake in the kernel function?

using CUDA
using AMDGPU
using KernelAbstractions

function compute_tensors(tensor, kernel_fun, Nx, Ny, Nz)
    kernel! = kernel_fun(get_backend(tensor), 512)
    kernel!(tensor, Nx, Ny, Nz; ndrange=size(tensor))
    KernelAbstractions.synchronize(get_backend(tensor))
    return nothing
end

@kernel function kernel_xx!(tensor, Nx::Int64, Ny::Int64, Nz::Int64)
    i, j, k = @index(Global, NTuple)
    sum = zero(eltype(tensor))
    for p in (-Nx):Nx, q in (-Ny):Ny
        sum += 2.0
    end
    @inbounds tensor[i, j, k] = sum
end

nx, ny, nz = 10, 1, 1
Nx, Ny, Nz = 1, 1, 1
tensor = zeros(Float64, nx, ny, nz)
compute_tensors(tensor, kernel_xx!, Nx, Ny, Nz)
println("cpu:", tensor)

tensor = CUDA.zeros(Float64, nx, ny, nz)
compute_tensors(tensor, kernel_xx!, Nx, Ny, Nz)
println("cuda:", tensor)

tensor = AMDGPU.zeros(Float64, nx, ny, nz)
compute_tensors(tensor, kernel_xx!, Nx, Ny, Nz)
println("amd:", tensor)
vchuravy commented 1 week ago

Can you also print the output?

ww1g11 commented 1 week ago

Can you also print the output?

The output is 

cpu: [18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0;;;]
cuda: [18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0;;;]
amd: [6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0;;;]
julia> versioninfo()
Julia Version 1.10.0
Commit 3120989f39 (2023-12-25 18:01 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: 12 × Intel(R) Core(TM) i7-9850H CPU @ 2.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, skylake)
  Threads: 1 on 12 virtual cores

[ Info: AMDGPU versioninfo
┌───────────┬──────────────────┬───────────┬───────────────────────────────────────────────────────────────────────────────────────────────┐
│ Available │ Name             │ Version   │ Path                                                                                          │
├───────────┼──────────────────┼───────────┼───────────────────────────────────────────────────────────────────────────────────────────────┤
│     +     │ LLD              │ -         │ C:\\Program Files\\AMD\\ROCm\\5.7\\bin\\ld.lld.exe                                            │
│     +     │ Device Libraries │ -         │ D:\\Softwares\\julia_pkg\\artifacts\\5ad5ecb46e3c334821f54c1feecc6c152b7b6a45\\amdgcn/bitcode │
│     +     │ HIP              │ 5.7.32000 │ C:\\WINDOWS\\SYSTEM32\\amdhip64.DLL                                                           │
│     +     │ rocBLAS          │ 3.1.0     │ C:\\Program Files\\AMD\\ROCm\\5.7\\bin\\rocblas.dll                                           │
│     +     │ rocSOLVER        │ 3.23.0    │ C:\\Program Files\\AMD\\ROCm\\5.7\\bin\\rocsolver.dll                                         │
│     +     │ rocALUTION       │ -         │ C:\\Program Files\\AMD\\ROCm\\5.7\\bin\\rocalution.dll                                        │
│     +     │ rocSPARSE        │ -         │ C:\\Program Files\\AMD\\ROCm\\5.7\\bin\\rocsparse.dll                                         │
│     +     │ rocRAND          │ 2.10.5    │ C:\\Program Files\\AMD\\ROCm\\5.7\\bin\\rocrand.dll                                           │
│     +     │ rocFFT           │ 1.0.27    │ C:\\Program Files\\AMD\\ROCm\\5.7\\bin\\rocfft.dll                                            │
│     -     │ MIOpen           │ -         │ -                                                                                             │
└───────────┴──────────────────┴───────────┴───────────────────────────────────────────────────────────────────────────────────────────────┘

[ Info: AMDGPU devices
┌────┬────────────────────────┬──────────┬───────────┬───────────┐
│ Id │                   Name │ GCN arch │ Wavefront │    Memory │
├────┼────────────────────────┼──────────┼───────────┼───────────┤
│  1 │ AMD Radeon RX 7600M XT │  gfx1102 │        32 │ 7.984 GiB │
└────┴────────────────────────┴──────────┴───────────┴───────────┘
vchuravy commented 1 week ago

Yeah that's concerning... @pxl-th any ideas?

luraess commented 1 week ago

AMDGPU on Windows right? Could one have sum differently to discern if p or q are not executed?

ww1g11 commented 1 week ago

An error occurs if the kernel function changes to 

@kernel function kernel_xx!(tensor, Nx::Int64, Ny::Int64, Nz::Int64)
    i, j, k = @index(Global, NTuple)
    sum = zero(eltype(tensor))
    for p in (-Nx):Nx
        for q in (-Ny):Ny
            sum += 2.0
        end
    end
    @inbounds tensor[i, j, k] = sum
end

the error msg:

julia> include("amd.jl")
cpu:[18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0;;;]
cuda:[18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0;;;]
ERROR: LoadError: Not implemented
Stacktrace:
  [1] error(s::String)
    @ Base .\error.jl:35
  [2] runtime_module(job::GPUCompiler.CompilerJob)
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\interface.jl:176
  [3] build_runtime(job::GPUCompiler.CompilerJob)
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\rtlib.jl:102
  [4] (::GPUCompiler.var"#140#142"{GPUCompiler.CompilerJob{…}})()
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\rtlib.jl:143
  [5] lock(f::GPUCompiler.var"#140#142"{GPUCompiler.CompilerJob{…}}, l::ReentrantLock)
    @ Base .\lock.jl:229
  [6] macro expansion
    @ D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\rtlib.jl:121 [inlined]
  [7] load_runtime(job::GPUCompiler.CompilerJob)
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\utils.jl:103
  [8] macro expansion
    @ D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:304 [inlined]
  [9]
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\utils.jl:103
 [10] emit_llvm
    @ D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\utils.jl:97 [inlined]
 [11]
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:136
 [12] codegen
    @ D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:115 [inlined]
 [13]
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:111
 [14] compile
    @ D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:103 [inlined]
 [15] #40
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\compiler\codegen.jl:170 [inlined]
 [16] JuliaContext(f::AMDGPU.Compiler.var"#40#41"{GPUCompiler.CompilerJob{…}}; kwargs::@Kwargs{})
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:52
 [17] JuliaContext(f::Function)
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:42
 [18] hipcompile(job::GPUCompiler.CompilerJob)
    @ AMDGPU.Compiler D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\compiler\codegen.jl:169
 [19] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(AMDGPU.Compiler.hipcompile), linker::typeof(AMDGPU.Compiler.hiplink))
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\execution.jl:128
 [20] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\execution.jl:103
 [21] macro expansion
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\compiler\codegen.jl:137 [inlined]
 [22] macro expansion
    @ .\lock.jl:267 [inlined]
 [23] hipfunction(f::GPUArrays.var"#6#7", tt::Type{Tuple{…}}; kwargs::@Kwargs{name::Nothing})
    @ AMDGPU.Compiler D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\compiler\codegen.jl:131
 [24] hipfunction
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\compiler\codegen.jl:130 [inlined]
 [25] macro expansion
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\highlevel.jl:172 [inlined]
 [26] #gpu_call#55
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\gpuarrays.jl:8 [inlined]
 [27] gpu_call
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\gpuarrays.jl:5 [inlined]
 [28] gpu_call(::GPUArrays.var"#6#7", ::ROCArray{…}, ::Float64; target::ROCArray{…}, elements::Nothing, threads::Nothing, blocks::Nothing, name::Nothing)
    @ GPUArrays D:\Softwares\julia_pkg\packages\GPUArrays\qt4ax\src\device\execution.jl:65
 [29] gpu_call
    @ GPUArrays D:\Softwares\julia_pkg\packages\GPUArrays\qt4ax\src\device\execution.jl:34 [inlined]
 [30] fill!(A::ROCArray{Float64, 3, AMDGPU.Runtime.Mem.HIPBuffer}, x::Float64)
    @ GPUArrays D:\Softwares\julia_pkg\packages\GPUArrays\qt4ax\src\host\construction.jl:14
 [31] zeros(::Type, ::Int64, ::Vararg{Int64})
    @ AMDGPU D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\array.jl:244
luraess commented 1 week ago

Testing on RX 7800 XT Ubuntu 22.04 I can reproduce the issue.

The following does not errors though for me:

@kernel function kernel_xx!(tensor, Nx::Int64, Ny::Int64, Nz::Int64)
    i, j, k = @index(Global, NTuple)
    sum = zero(eltype(tensor))
    for p in (-Nx):Nx
        for q in (-Ny):Ny
            sum += 2.0
        end
    end
    @inbounds tensor[i, j, k] = sum
end

and produces the same wrong results as the original version.

From testing, it seems that -Ny brakes the thing as following produces correct result:

@kernel function kernel_xx!(tensor, Nx::Int64, Ny::Int64, Nz::Int64)
    i, j, k = @index(Global, NTuple)
    sum = zero(eltype(tensor))
    for p in (-Nx):Nx
        for q in (-1):Ny
            sum += 2.0
        end
    end
    @inbounds tensor[i, j, k] = sum
end

EDIT: (-Ny):Ny seems to be the deal-breaker in both loop-variants.

luraess commented 1 week ago

Also the "plain" AMDGPU version works fine:

using AMDGPU

function compute_amdgpu(tensor, kernel_fun, Nx, Ny, Nz)
    groupsize = (16, 4, 2)                     # nthreads
    gridsize  = cld.(size(tensor), groupsize)  # nblocks
    @roc groupsize=groupsize gridsize=gridsize kernel_fun(tensor, Nx, Ny, Nz)
    AMDGPU.synchronize()
    return nothing
end

function kernel_xx2!(tensor, Nx, Ny, Nz)
    i = (workgroupIdx().x - 1) * workgroupDim().x + workitemIdx().x
    j = (workgroupIdx().y - 1) * workgroupDim().y + workitemIdx().y
    k = (workgroupIdx().z - 1) * workgroupDim().z + workitemIdx().z
    sum = zero(eltype(tensor))
    for p in (-Nx):Nx, q in (-Ny):Ny
        sum += 2.0
    end
    if i ∈ axes(tensor, 1) && j ∈ axes(tensor, 2) && k ∈ axes(tensor, 3)
        @inbounds tensor[i, j, k] = sum
    end
    return nothing
end

nx, ny, nz = 10, 1, 1
Nx, Ny, Nz = 1, 1, 1

tensor2 = AMDGPU.zeros(Float64, nx, ny, nz)
compute_amdgpu(tensor2, kernel_xx2!, Nx, Ny, Nz)
println("amdgpu:", tensor2)
pxl-th commented 1 week ago

This seems to appear only when using NTuple index type. Changing to linear or cartesian works fine:

@kernel function kernel_xx!(tensor, Nx::Int64, Ny::Int64, Nz::Int64)
    i = @index(Global)
    s = zero(eltype(tensor))
    for p in (-Nx):Nx, q in (-Ny):Ny
        s += 2.0
    end
    @inbounds tensor[i] = s
end
pxl-th commented 1 week ago

Or more generally, when passing size(x) to ndrage instead of length(x).

pxl-th commented 1 week ago

Here's optimized LLVM IR for:

@kernel function kernel_xx!(tensor, Nx::Int64, Ny::Int64, Nz::Int64)
    idx = @index(Global)
    res = zero(eltype(tensor))
    for p in (-Nx):Nx
        for q in (-Ny):Ny
            res += 2.0
        end
    end
    @inbounds tensor[idx] = res
end

@vchuravy in size(x) case, there's lot's of comparisons >= 0 and terminations in case of false. I suspect that's why it fails:

%29 = icmp sgt i64 %.fca.1.1.0.1.0.extract, 0
br i1 %29, label %pass6, label %fail5

; Function Attrs: nounwind readnone speculatable willreturn declare i32 @llvm.amdgcn.workgroup.id.x() #0

; Function Attrs: nounwind readnone speculatable willreturn declare i32 @llvm.amdgcn.workitem.id.x() #0

; Function Attrs: cold noreturn nounwind declare void @llvm.amdgcn.endpgm() #1

; Function Attrs: nocallback nofree nosync nounwind readnone speculatable willreturn declare i64 @llvm.smax.i64(i64, i64) #2

define amdgpu_kernel void @_Z14gpu_kernel_xx_16CompilerMetadataI11DynamicSize12DynamicCheckv16CartesianIndicesILl3E5TupleI5OneToI5Int64ES4_IS5_ES4_IS5_EEE7NDRangeILl3ES0_S0_S2_ILl3ES3_IS4_IS5_ES4_IS5_ES4_IS5_EEES2_ILl3ES3_IS4_IS5_ES4_IS5_ES4_IS5_EEEEE14ROCDeviceArrayI7Float64Ll3ELl1EES5_S5S5({ i64, i64, i64, i64, i64, i64, i32, i32, i64, i64, i64, i64 } %state, { [1 x [3 x [1 x i64]]], [2 x [1 x [3 x [1 x i64]]]] } %0, { [3 x i64], i8 addrspace(1), i64 } %1, i64 signext %2, i64 signext %3, i64 signext %4) local_unnamed_addr #3 { conversion: %.fca.0.0.0.0.extract = extractvalue { [1 x [3 x [1 x i64]]], [2 x [1 x [3 x [1 x i64]]]] } %0, 0, 0, 0, 0 %.fca.0.0.1.0.extract = extractvalue { [1 x [3 x [1 x i64]]], [2 x [1 x [3 x [1 x i64]]]] } %0, 0, 0, 1, 0 %.fca.0.0.2.0.extract = extractvalue { [1 x [3 x [1 x i64]]], [2 x [1 x [3 x [1 x i64]]]] } %0, 0, 0, 2, 0 %.fca.1.0.0.0.0.extract = extractvalue { [1 x [3 x [1 x i64]]], [2 x [1 x [3 x [1 x i64]]]] } %0, 1, 0, 0, 0, 0 %.fca.1.0.0.1.0.extract = extractvalue { [1 x [3 x [1 x i64]]], [2 x [1 x [3 x [1 x i64]]]] } %0, 1, 0, 0, 1, 0 %.fca.1.1.0.0.0.extract = extractvalue { [1 x [3 x [1 x i64]]], [2 x [1 x [3 x [1 x i64]]]] } %0, 1, 1, 0, 0, 0 %.fca.1.1.0.1.0.extract = extractvalue { [1 x [3 x [1 x i64]]], [2 x [1 x [3 x [1 x i64]]]] } %0, 1, 1, 0, 1, 0 %.fca.1.1.0.2.0.extract = extractvalue { [1 x [3 x [1 x i64]]], [2 x [1 x [3 x [1 x i64]]]] } %0, 1, 1, 0, 2, 0 %.fca.1.extract = extractvalue { [3 x i64], i8 addrspace(1), i64 } %1, 1 %5 = call i32 @llvm.amdgcn.workitem.id.x(), !range !7 %6 = call i64 @llvm.smax.i64(i64 %.fca.1.0.0.1.0.extract, i64 0) %7 = icmp sgt i64 %.fca.1.0.0.0.0.extract, 0 br i1 %7, label %pass, label %fail

L674: ; preds = %L674.preheader, %L706 %value_phi17 = phi i64 [ %10, %L706 ], [ %65, %L674.preheader ] %value_phi18 = phi double [ %value_phi26, %L706 ], [ 0.000000e+00, %L674.preheader ] br i1 %.not127.not, label %L706, label %L693

L693: ; preds = %L693, %L674 %value_phi22 = phi double [ %8, %L693 ], [ %value_phi18, %L674 ] %value_phi23 = phi i64 [ %9, %L693 ], [ %67, %L674 ] %8 = fadd double %value_phi22, 2.000000e+00 %.not128 = icmp eq i64 %value_phi23, %value_phi19 %9 = add i64 %value_phi23, 1 br i1 %.not128, label %L706, label %L693

L706: ; preds = %L693, %L674 %value_phi26 = phi double [ %value_phi18, %L674 ], [ %8, %L693 ] %.not129 = icmp eq i64 %value_phi17, %value_phi %10 = add i64 %value_phi17, 1 br i1 %.not129, label %L732, label %L674

L732: ; preds = %pass10, %L706 %value_phi29 = phi double [ 0.000000e+00, %pass10 ], [ %value_phi26, %L706 ] %11 = add i64 %64, %32 %reass.add137 = add i64 %11, %reass.mul135 %reass.mul138 = mul i64 %reass.add137, %60 %12 = add i64 %reass.mul138, %31 %13 = add i64 %12, %reass.mul %14 = bitcast i8 addrspace(1) %.fca.1.extract to double addrspace(1) %15 = getelementptr inbounds double, double addrspace(1) %14, i64 %13 store double %value_phi29, double addrspace(1) %15, align 8, !tbaa !8 br label %L738

L738: ; preds = %pass6, %L732 ret void

fail: ; preds = %conversion %state.i.fca.0.extract.i = extractvalue { i64, i64, i64, i64, i64, i64, i32, i32, i64, i64, i64, i64 } %state, 0 %16 = inttoptr i64 %state.i.fca.0.extract.i to i32 store i32 1, i32 %16, align 1 call void @llvm.amdgcn.endpgm() unreachable

pass: ; preds = %conversion %17 = call i32 @llvm.amdgcn.workgroup.id.x(), !range !11 %18 = zext i32 %17 to i64 %19 = udiv i64 %18, %.fca.1.0.0.0.0.extract %20 = mul i64 %19, %.fca.1.0.0.0.0.extract %21 = icmp sgt i64 %.fca.1.0.0.1.0.extract, 0 br i1 %21, label %pass2, label %fail1

fail1: ; preds = %pass %state.i.fca.0.extract.i28 = extractvalue { i64, i64, i64, i64, i64, i64, i32, i32, i64, i64, i64, i64 } %state, 0 %22 = inttoptr i64 %state.i.fca.0.extract.i28 to i32 store i32 1, i32 %22, align 1 call void @llvm.amdgcn.endpgm() unreachable

pass2: ; preds = %pass %23 = udiv i64 %19, %6 %24 = mul i64 %23, %6 %25 = sub i64 %19, %24 %26 = call i64 @llvm.smax.i64(i64 %.fca.1.1.0.1.0.extract, i64 0) %27 = icmp sgt i64 %.fca.1.1.0.0.0.extract, 0 br i1 %27, label %pass4, label %fail3

fail3: ; preds = %pass2 %state.i.fca.0.extract.i42 = extractvalue { i64, i64, i64, i64, i64, i64, i32, i32, i64, i64, i64, i64 } %state, 0 %28 = inttoptr i64 %state.i.fca.0.extract.i42 to i32 store i32 1, i32 %28, align 1 call void @llvm.amdgcn.endpgm() unreachable

pass4: ; preds = %pass2 %29 = icmp sgt i64 %.fca.1.1.0.1.0.extract, 0 br i1 %29, label %pass6, label %fail5

fail5: ; preds = %pass4 %state.i.fca.0.extract.i56 = extractvalue { i64, i64, i64, i64, i64, i64, i32, i32, i64, i64, i64, i64 } %state, 0 %30 = inttoptr i64 %state.i.fca.0.extract.i56 to i32 store i32 1, i32 %30, align 1 call void @llvm.amdgcn.endpgm() unreachable

pass6: ; preds = %pass4 %31 = zext i32 %5 to i64 %32 = udiv i64 %31, %.fca.1.1.0.0.0.extract %33 = udiv i64 %32, %26 %34 = mul i64 %33, %26 %35 = add i64 %20, %32 %reass.add = sub i64 %18, %35 %reass.mul = mul i64 %reass.add, %.fca.1.1.0.0.0.extract %36 = add nuw nsw i64 %31, 1 %37 = add i64 %36, %reass.mul %38 = mul i64 %25, %.fca.1.1.0.1.0.extract %39 = add i64 %38, 1 %40 = add i64 %39, %32 %41 = sub i64 %40, %34 %42 = mul i64 %23, %.fca.1.1.0.2.0.extract %43 = add i64 %42, 1 %44 = add i64 %43, %33 %45 = icmp sgt i64 %37, 0 %46 = icmp sle i64 %37, %.fca.0.0.0.0.extract %47 = and i1 %45, %46 %48 = icmp sgt i64 %41, 0 %49 = icmp sle i64 %41, %.fca.0.0.1.0.extract %50 = and i1 %48, %49 %51 = icmp sgt i64 %44, 0 %52 = icmp sle i64 %44, %.fca.0.0.2.0.extract %53 = and i1 %51, %52 %54 = and i1 %47, %50 %55 = and i1 %53, %54 br i1 %55, label %pass10, label %L738

pass10: ; preds = %pass6 %56 = udiv i64 %19, %.fca.1.0.0.1.0.extract %57 = mul i64 %56, %.fca.1.0.0.1.0.extract %58 = udiv i64 %32, %.fca.1.1.0.1.0.extract %59 = mul i64 %56, %.fca.1.1.0.2.0.extract %60 = call i64 @llvm.smax.i64(i64 %.fca.0.0.0.0.extract, i64 0) %61 = call i64 @llvm.smax.i64(i64 %.fca.0.0.1.0.extract, i64 0) %62 = add i64 %57, %58 %reass.add134 = sub i64 %19, %62 %reass.mul135 = mul i64 %reass.add134, %.fca.1.1.0.1.0.extract %63 = add i64 %58, %59 %64 = mul i64 %63, %61 %65 = sub i64 0, %2 %.not = icmp sgt i64 %65, %2 %66 = sext i1 %.not to i64 %value_phi = xor i64 %66, %2 %.not125.not = icmp slt i64 %value_phi, %65 br i1 %.not125.not, label %L732, label %L674.preheader

L674.preheader: ; preds = %pass10 %67 = sub i64 0, %3 %.not126 = icmp sgt i64 %67, %3 %68 = sext i1 %.not126 to i64 %value_phi19 = xor i64 %68, %3 %.not127.not = icmp slt i64 %value_phi19, %67 br label %L674 }

attributes #0 = { nounwind readnone speculatable willreturn } attributes #1 = { cold noreturn nounwind } attributes #2 = { nocallback nofree nosync nounwind readnone speculatable willreturn } attributes #3 = { "amdgpu-unsafe-fp-atomics"="true" "target-cpu"="gfx1100" "target-features"="+wavefrontsize32,-wavefrontsize64" }

!llvm.module.flags = !{!0, !1, !2, !3} !opencl.ocl.version = !{!4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4} !llvm.ident = !{!5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5} !julia.kernel = !{!6}

!0 = !{i32 2, !"Dwarf Version", i32 4} !1 = !{i32 2, !"Debug Info Version", i32 3} !2 = !{i32 1, !"wchar_size", i32 4} !3 = !{i32 7, !"PIC Level", i32 1} !4 = !{i32 2, i32 0} !5 = !{!"clang version 15.0.0 (/cache/yggdrasil/downloads/clones/llvm-project.git-974efd367bc513231526d317489c66cb27727ef3caa41108e3819c131a8acf57 f3d695fc2985a8dfdd5f4219d351fdeac3038867)"} !6 = !{void ({ i64, i64, i64, i64, i64, i64, i32, i32, i64, i64, i64, i64 }, { [1 x [3 x [1 x i64]]], [2 x [1 x [3 x [1 x i64]]]] }, { [3 x i64], i8 addrspace(1), i64 }, i64, i64, i64) @_Z14gpu_kernel_xx_16CompilerMetadataI11DynamicSize12DynamicCheckv16CartesianIndicesILl3E5TupleI5OneToI5Int64ES4_IS5_ES4_IS5_EEE7NDRangeILl3ES0_S0_S2_ILl3ES3_IS4_IS5_ES4_IS5_ES4_IS5_EEES2_ILl3ES3_IS4_IS5_ES4_IS5_ES4_IS5_EEEEE14ROCDeviceArrayI7Float64Ll3ELl1EES5_S5S5} !7 = !{i32 0, i32 1023} !8 = !{!9, !9, i64 0, i64 0} !9 = !{!"custom_tbaa_addrspace(1)", !10, i64 0} !10 = !{!"custom_tbaa"} !11 = !{i32 0, i32 -2}

- `ndrange=length(x)`:
```llvm
; ModuleID = 'start'
source_filename = "start"
target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:10:11:12:13"
target triple = "amdgcn-amd-amdhsa"

; Function Attrs: nounwind readnone speculatable willreturn
declare i32 @llvm.amdgcn.workgroup.id.x() #0

; Function Attrs: nounwind readnone speculatable willreturn
declare i32 @llvm.amdgcn.workitem.id.x() #0

define amdgpu_kernel void @_Z14gpu_kernel_xx_16CompilerMetadataI11DynamicSize12DynamicCheckv16CartesianIndicesILl1E5TupleI5OneToI5Int64EEE7NDRangeILl1ES0_S0_S2_ILl1ES3_IS4_IS5_EEES2_ILl1ES3_IS4_IS5_EEEEE14ROCDeviceArrayI7Float64Ll3ELl1EES5_S5_S5_({ i64, i64, i64, i64, i64, i64, i32, i32, i64, i64, i64, i64 } %state, { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %0, { [3 x i64], i8 addrspace(1)*, i64 } %1, i64 signext %2, i64 signext %3, i64 signext %4) local_unnamed_addr #1 {
conversion:
  %.fca.0.0.0.0.extract = extractvalue { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %0, 0, 0, 0, 0
  %.fca.1.1.0.0.0.extract = extractvalue { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %0, 1, 1, 0, 0, 0
  %.fca.1.extract = extractvalue { [3 x i64], i8 addrspace(1)*, i64 } %1, 1
  %5 = call i32 @llvm.amdgcn.workgroup.id.x(), !range !7
  %6 = call i32 @llvm.amdgcn.workitem.id.x(), !range !8
  %7 = add nuw nsw i32 %6, 1
  %8 = zext i32 %7 to i64
  %9 = zext i32 %5 to i64
  %10 = mul i64 %.fca.1.1.0.0.0.extract, %9
  %11 = add i64 %10, %8
  %12 = icmp slt i64 %11, 1
  %13 = icmp sgt i64 %11, %.fca.0.0.0.0.extract
  %14 = or i1 %12, %13
  br i1 %14, label %L299, label %L103

L103:                                             ; preds = %conversion
  %15 = sub i64 0, %2
  %.not = icmp sgt i64 %15, %2
  %16 = sext i1 %.not to i64
  %value_phi = xor i64 %16, %2
  %.not6.not = icmp slt i64 %value_phi, %15
  br i1 %.not6.not, label %L293, label %L235.preheader

L235.preheader:                                   ; preds = %L103
  %17 = sub i64 0, %3
  %.not7 = icmp sgt i64 %17, %3
  %18 = sext i1 %.not7 to i64
  %value_phi5 = xor i64 %18, %3
  %.not8.not = icmp slt i64 %value_phi5, %17
  br label %L235

L235:                                             ; preds = %L267, %L235.preheader
  %value_phi3 = phi i64 [ %21, %L267 ], [ %15, %L235.preheader ]
  %value_phi4 = phi double [ %value_phi12, %L267 ], [ 0.000000e+00, %L235.preheader ]
  br i1 %.not8.not, label %L267, label %L254

L254:                                             ; preds = %L254, %L235
  %value_phi8 = phi double [ %19, %L254 ], [ %value_phi4, %L235 ]
  %value_phi9 = phi i64 [ %20, %L254 ], [ %17, %L235 ]
  %19 = fadd double %value_phi8, 2.000000e+00
  %.not9 = icmp eq i64 %value_phi9, %value_phi5
  %20 = add i64 %value_phi9, 1
  br i1 %.not9, label %L267, label %L254

L267:                                             ; preds = %L254, %L235
  %value_phi12 = phi double [ %value_phi4, %L235 ], [ %19, %L254 ]
  %.not10 = icmp eq i64 %value_phi3, %value_phi
  %21 = add i64 %value_phi3, 1
  br i1 %.not10, label %L293, label %L235

L293:                                             ; preds = %L267, %L103
  %value_phi15 = phi double [ 0.000000e+00, %L103 ], [ %value_phi12, %L267 ]
  %22 = add nsw i64 %8, -1
  %23 = add i64 %22, %10
  %24 = bitcast i8 addrspace(1)* %.fca.1.extract to double addrspace(1)*
  %25 = getelementptr inbounds double, double addrspace(1)* %24, i64 %23
  store double %value_phi15, double addrspace(1)* %25, align 8, !tbaa !9
  br label %L299

L299:                                             ; preds = %L293, %conversion
  ret void
}

attributes #0 = { nounwind readnone speculatable willreturn }
attributes #1 = { "amdgpu-unsafe-fp-atomics"="true" "target-cpu"="gfx1100" "target-features"="+wavefrontsize32,-wavefrontsize64" }

!llvm.module.flags = !{!0, !1, !2, !3}
!opencl.ocl.version = !{!4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4, !4}
!llvm.ident = !{!5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5, !5}
!julia.kernel = !{!6}

!0 = !{i32 2, !"Dwarf Version", i32 4}
!1 = !{i32 2, !"Debug Info Version", i32 3}
!2 = !{i32 1, !"wchar_size", i32 4}
!3 = !{i32 7, !"PIC Level", i32 1}
!4 = !{i32 2, i32 0}
!5 = !{!"clang version 15.0.0 (/cache/yggdrasil/downloads/clones/llvm-project.git-974efd367bc513231526d317489c66cb27727ef3caa41108e3819c131a8acf57 f3d695fc2985a8dfdd5f4219d351fdeac3038867)"}
!6 = !{void ({ i64, i64, i64, i64, i64, i64, i32, i32, i64, i64, i64, i64 }, { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] }, { [3 x i64], i8 addrspace(1)*, i64 }, i64, i64, i64)* @_Z14gpu_kernel_xx_16CompilerMetadataI11DynamicSize12DynamicCheckv16CartesianIndicesILl1E5TupleI5OneToI5Int64EEE7NDRangeILl1ES0_S0_S2_ILl1ES3_IS4_IS5_EEES2_ILl1ES3_IS4_IS5_EEEEE14ROCDeviceArrayI7Float64Ll3ELl1EES5_S5_S5_}
!7 = !{i32 0, i32 -2}
!8 = !{i32 0, i32 1023}
!9 = !{!10, !10, i64 0, i64 0}
!10 = !{!"custom_tbaa_addrspace(1)", !11, i64 0}
!11 = !{!"custom_tbaa"}
ww1g11 commented 6 days ago

In my computer, the linear or cartesian also give the wrong results:

cpu:[18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0;;;]
amd:[6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0; 6.0;;;]

This seems to appear only when using NTuple index type. Changing to linear or cartesian works fine:

@kernel function kernel_xx!(tensor, Nx::Int64, Ny::Int64, Nz::Int64)
    i = @index(Global)
    s = zero(eltype(tensor))
    for p in (-Nx):Nx, q in (-Ny):Ny
        s += 2.0
    end
    @inbounds tensor[i] = s
end
ww1g11 commented 6 days ago

This time this kernel works well (the results are still wrong).

An error occurs if the kernel function changes to

@kernel function kernel_xx!(tensor, Nx::Int64, Ny::Int64, Nz::Int64)
    i, j, k = @index(Global, NTuple)
    sum = zero(eltype(tensor))
    for p in (-Nx):Nx
        for q in (-Ny):Ny
            sum += 2.0
        end
    end
    @inbounds tensor[i, j, k] = sum
end

the error msg:

julia> include("amd.jl")
cpu:[18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0;;;]
cuda:[18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0; 18.0;;;]
ERROR: LoadError: Not implemented
Stacktrace:
  [1] error(s::String)
    @ Base .\error.jl:35
  [2] runtime_module(job::GPUCompiler.CompilerJob)
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\interface.jl:176
  [3] build_runtime(job::GPUCompiler.CompilerJob)
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\rtlib.jl:102
  [4] (::GPUCompiler.var"#140#142"{GPUCompiler.CompilerJob{…}})()
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\rtlib.jl:143
  [5] lock(f::GPUCompiler.var"#140#142"{GPUCompiler.CompilerJob{…}}, l::ReentrantLock)
    @ Base .\lock.jl:229
  [6] macro expansion
    @ D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\rtlib.jl:121 [inlined]
  [7] load_runtime(job::GPUCompiler.CompilerJob)
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\utils.jl:103
  [8] macro expansion
    @ D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:304 [inlined]
  [9]
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\utils.jl:103
 [10] emit_llvm
    @ D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\utils.jl:97 [inlined]
 [11]
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:136
 [12] codegen
    @ D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:115 [inlined]
 [13]
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:111
 [14] compile
    @ D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:103 [inlined]
 [15] #40
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\compiler\codegen.jl:170 [inlined]
 [16] JuliaContext(f::AMDGPU.Compiler.var"#40#41"{GPUCompiler.CompilerJob{…}}; kwargs::@Kwargs{})
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:52
 [17] JuliaContext(f::Function)
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\driver.jl:42
 [18] hipcompile(job::GPUCompiler.CompilerJob)
    @ AMDGPU.Compiler D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\compiler\codegen.jl:169
 [19] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(AMDGPU.Compiler.hipcompile), linker::typeof(AMDGPU.Compiler.hiplink))
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\execution.jl:128
 [20] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler D:\Softwares\julia_pkg\packages\GPUCompiler\nWT2N\src\execution.jl:103
 [21] macro expansion
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\compiler\codegen.jl:137 [inlined]
 [22] macro expansion
    @ .\lock.jl:267 [inlined]
 [23] hipfunction(f::GPUArrays.var"#6#7", tt::Type{Tuple{…}}; kwargs::@Kwargs{name::Nothing})
    @ AMDGPU.Compiler D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\compiler\codegen.jl:131
 [24] hipfunction
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\compiler\codegen.jl:130 [inlined]
 [25] macro expansion
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\highlevel.jl:172 [inlined]
 [26] #gpu_call#55
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\gpuarrays.jl:8 [inlined]
 [27] gpu_call
    @ D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\gpuarrays.jl:5 [inlined]
 [28] gpu_call(::GPUArrays.var"#6#7", ::ROCArray{…}, ::Float64; target::ROCArray{…}, elements::Nothing, threads::Nothing, blocks::Nothing, name::Nothing)
    @ GPUArrays D:\Softwares\julia_pkg\packages\GPUArrays\qt4ax\src\device\execution.jl:65
 [29] gpu_call
    @ GPUArrays D:\Softwares\julia_pkg\packages\GPUArrays\qt4ax\src\device\execution.jl:34 [inlined]
 [30] fill!(A::ROCArray{Float64, 3, AMDGPU.Runtime.Mem.HIPBuffer}, x::Float64)
    @ GPUArrays D:\Softwares\julia_pkg\packages\GPUArrays\qt4ax\src\host\construction.jl:14
 [31] zeros(::Type, ::Int64, ::Vararg{Int64})
    @ AMDGPU D:\Softwares\julia_pkg\packages\AMDGPU\WqMSe\src\array.jl:244
pxl-th commented 6 days ago

In my computer, the linear or cartesian also give the wrong results:

Did you change how you launch the code? You need to specity ndrange=length(x) instead of ndrange=size(x). Because that's the main issue, passing a tuple to ndrange.

ww1g11 commented 3 days ago

In my computer, the linear or cartesian also give the wrong results:

Did you change how you launch the code? You need to specity ndrange=length(x) instead of ndrange=size(x). Because that's the main issue, passing a tuple to ndrange.

I just tried ndrange=length(x), and it give the correct results.

I am using ndrange=size(x) because I want to know its Cartesian Indices, it seems I can cast the linear index to CartesianIndices manually: 

    idx = @index(Global)
    i, j = Tuple(CartesianIndices(tensor)[idx])

so using ndrange=length(x) solves the problem.

pxl-th commented 3 days ago

You can also access size of the array within the kernel and compute i, j indices from there. And if you do only element-wise operations, you can just index with idx since it will iterate all elements of x anyway:

x[2, 2] == x[4]
ww1g11 commented 3 days ago

You can also access size of the array within the kernel and compute i, j indices from there. And if you do only element-wise operations, you can just index with idx since it will iterate all elements of x anyway:

x[2, 2] == x[4]

thanks~

vchuravy commented 3 days ago

So this is smelling more and more like a compiler bug...

The below code is an attempt of mine to remove the KA syntax sugar. Next step is the inline expand and then we should have a KA free reproducer.

import KernelAbstractions as KA
function gpu_kernel_xx!(ndrange, iterspace, tensor, Nx::Int64, Ny::Int64; )
    I = @inbounds KA.expand(iterspace, AMDGPU.blockIdx().x, AMDGPU.threadIdx().x)
    if I in ndrange
        (i, j, k) = @inbounds KA.expand(iterspace, AMDGPU.blockIdx().x, AMDGPU.threadIdx().x).I
        sum = zero(eltype(tensor))
        for p = 1:Nx + 2
            for q = -Ny:Ny
                sum += 1.0
            end
        end
        @inbounds tensor[i, j, k] = sum
    end
    return nothing
end

obj = kernel_xx!(ROCBackend())
ndrange, wgsize, iterspace, dynamic = KernelAbstractions.launch_config(obj, (10, 1, 1), nothing)
ctx = KernelAbstractions.mkcontext(obj, ndrange, iterspace)

tensor = AMDGPU.zeros(Float64, nx, ny, nz)
@roc groupsize=512 gridsize=size(tensor) gpu_kernel_xx!(CartesianIndices((10, 1, 1)), iterspace, tensor, Nx, Ny)
println("ka_direct:", tensor)
vchuravy commented 3 days ago

This is now KA free:

"""
    assume(cond::Bool)

Assume that the condition `cond` is true. This is a hint to the compiler, possibly enabling
it to optimize more aggressively.
"""
@inline assume(cond::Bool) = Base.llvmcall(("""
        declare void @llvm.assume(i1)

        define void @entry(i8) #0 {
            %cond = icmp eq i8 %0, 1
            call void @llvm.assume(i1 %cond)
            ret void
        }

        attributes #0 = { alwaysinline }""", "entry"),
    Nothing, Tuple{Bool}, cond)

@inline function assume_nonzero(CI::CartesianIndices)
    ntuple(Val(ndims(CI))) do I
        @inline
        indices = CI.indices[I]
        assume(indices.stop > 0)
    end
end

Base.@propagate_inbounds function expand(blocks, workitems, groupidx::Integer, idx::Integer)
    # this causes a exception branch and a div
    assume_nonzero(blocks)
    assume_nonzero(workitems)
    expand(blocks, workitems, blocks[groupidx], workitems[idx])
end

@inline function expand(blocks, workitems, groupidx::CartesianIndex{N}, idx::CartesianIndex{N}) where {N}
    nI = ntuple(Val(N)) do I
        Base.@_inline_meta
        stride = size(workitems, I)
        gidx = groupidx.I[I]
        (gidx - 1) * stride + idx.I[I]
    end
    CartesianIndex(nI)
end

function gpu_kernel_xx!(ndrange, blocks, workitems, tensor, Nx::Int64, Ny::Int64; )
    I = @inbounds expand(blocks, workitems, AMDGPU.blockIdx().x, AMDGPU.threadIdx().x)
    if I in ndrange
        (i, j, k) = @inbounds expand(blocks, workitems, AMDGPU.blockIdx().x, AMDGPU.threadIdx().x).I
        sum = zero(eltype(tensor))
        for p = 1:Nx + 2
            for q = -Ny:Ny
                sum += 1.0
            end
        end
        @inbounds tensor[i, j, k] = sum
    end
    return nothing
end

tensor = AMDGPU.zeros(Float64, nx, ny, nz)

ndrange = CartesianIndices((10,1,1))
blocks = CartesianIndices((1, 1, 1))
workitems = CartesianIndices((10, 1, 1))

@roc groupsize=512 gridsize=size(tensor) gpu_kernel_xx!(ndrange, blocks, workitems, tensor, Nx, Ny)
println("ka_direct:", tensor)