Oddities with loop with dummy variable in Tullio

General Usage
, ,
1

Hi there

I spotred possibly a strange bug using Tullio on CPU. Here is my package status (Julia-1.8.2)

[bdcacae8] LoopVectorization v0.12.141
[bc48ee85] Tullio v0.3.5

And my minimal non-wroking example

using LoopVectorization, Tullio
function mytest_tullio(wdelta,probnew)
    q = size(wdelta,3)
    @tullio mat[a,b,j] := wdelta[j,m,a]*probnew[m,a] (a in 1:q, b in 1:q)
end

to generate the error (at least on a linux machine see below for more info)

julia> a = rand(2,3,4); b= rand(3,4); mytest_tullio(a,b)

here below the output

raised error 
ERROR: MethodError: no method matching reduced_add(::Float64, ::VectorizationBase.Vec{8, Float64})
Closest candidates are:
  reduced_add(::Union{Bool, Float16, Float32, Float64, Int16, Int32, Int64, Int8, UInt16, UInt32, UInt64, UInt8, SIMDTypes.Bit}, ::Union{Bool, Float16, Float32, Float64, Int16, Int32, Int64, Int8, UInt16, UInt32, UInt64, UInt8, SIMDTypes.Bit}) at ~/.julia/packages/VectorizationBase/e4FnQ/src/llvm_intrin/intrin_funcs.jl:690
  reduced_add(::VectorizationBase.AbstractSIMD{W}, ::VectorizationBase.AbstractSIMD{W}) where W at ~/.julia/packages/VectorizationBase/e4FnQ/src/llvm_intrin/intrin_funcs.jl:691
  reduced_add(::VectorizationBase.AbstractSIMD, ::VectorizationBase.AbstractSIMD) at ~/.julia/packages/VectorizationBase/e4FnQ/src/llvm_intrin/intrin_funcs.jl:692
Stacktrace:
  [1] macro expansion
    @ ~/.julia/packages/LoopVectorization/DDH6Z/src/reconstruct_loopset.jl:965 [inlined]
  [2] _turbo_!(::Val{(false, 0, 0, 0, false, 8, 64, 32, 64, 0x0000000000000001, 1, true)}, ::Val{(Symbol("##DROPPED#CONSTANT##"), Symbol("##DROPPED#CONSTANT##"), LoopVectorization.OperationStruct(0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.constant, 0x0001, 0x00), Symbol("##GLOBAL##CONSTANT##"), :nothing, LoopVectorization.OperationStruct(0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.constant, 0x0002, 0x00), Symbol("##DROPPED#CONSTANT##"), Symbol("##DROPPED#CONSTANT##"), LoopVectorization.OperationStruct(0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.constant, 0x0003, 0x00), :numericconstant, Symbol("###zero###18###"), LoopVectorization.OperationStruct(0x00000000000000000000000000000123, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.constant, 0x0004, 0x00), :LoopVectorization, :conditionalload, LoopVectorization.OperationStruct(0x00000000000000000000000000000321, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000006, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.memload, 0x0005, 0x01), :LoopVectorization, :LOOPCONSTANTINSTRUCTION, LoopVectorization.OperationStruct(0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.constant, 0x0006, 0x00), :LoopVectorization, :getindex, LoopVectorization.OperationStruct(0x00000000000000000000000000000143, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.memload, 0x0007, 0x02), :LoopVectorization, :getindex, LoopVectorization.OperationStruct(0x00000000000000000000000000000043, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.memload, 0x0008, 0x03), :numericconstant, Symbol("###reduction##zero###29###"), LoopVectorization.OperationStruct(0x00000000000000000000000000000321, 0x00000000000000000000000000000000, 0x00000000000000000000000000000004, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.constant, 0x0009, 0x00), :LoopVectorization, :vfmadd_fast, LoopVectorization.OperationStruct(0x00000000000000000000000000000143, 0x00000000000000000000000000000004, 0x00000000000000000000000000000000, 0x00000000000000000000000700080009, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.compute, 0x0009, 0x00), :LoopVectorization, :reduced_add, LoopVectorization.OperationStruct(0x00000000000000000000000000000321, 0x00000000000000000000000000000004, 0x00000000000000000000000000000000, 0x000000000000000000000000000a0005, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.compute, 0x0005, 0x00), :LoopVectorization, :setindex!, LoopVectorization.OperationStruct(0x00000000000000000000000000000321, 0x00000000000000000000000000000004, 0x00000000000000000000000000000000, 0x0000000000000000000000000000000b, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, LoopVectorization.memstore, 0x000a, 0x01))}, ::Val{(LoopVectorization.ArrayRefStruct{:β„›, Symbol("##vptr##_β„›")}(0x00000000000000000000000000010101, 0x00000000000000000000000000030201, 0x00000000000000000000000000000000, 0x00000000000000000000000000010101), LoopVectorization.ArrayRefStruct{:wdelta, Symbol("##vptr##_wdelta")}(0x00000000000000000000000000010101, 0x00000000000000000000000000010403, 0x00000000000000000000000000000000, 0x00000000000000000000000000010101), LoopVectorization.ArrayRefStruct{:probnew, Symbol("##vptr##_probnew")}(0x00000000000000000000000000000101, 0x00000000000000000000000000000403, 0x00000000000000000000000000000000, 0x00000000000000000000000000000101))}, ::Val{(0, (), (1, 3, 6), (), (), ((4, LoopVectorization.IntOrFloat), (9, LoopVectorization.IntOrFloat)), ())}, ::Val{(:j, :b, :a, :m)}, ::Val{Tuple{NTuple{4, Static.OptionallyStaticUnitRange{Static.StaticInt{0}, Int64}}, Tuple{LayoutPointers.GroupedStridedPointers{Tuple{Ptr{Float64}, Ptr{Float64}, Ptr{Float64}}, (1, 1, 1), (0, 0, 0), ((1, 2, 3), (1, 2, 3), (1, 2)), ((1, 2, 3), (4, 5, 6), (7, 8)), Tuple{Static.StaticInt{8}, Int64, Int64, Static.StaticInt{8}, Int64, Int64, Static.StaticInt{8}, Int64}, NTuple{8, Static.StaticInt{0}}}, Bool}}}, ::Int64, ::Int64, ::Int64, ::Int64, ::Ptr{Float64}, ::Ptr{Float64}, ::Ptr{Float64}, ::Int64, ::Int64, ::Int64, ::Int64, ::Int64, ::Bool)
    @ LoopVectorization ~/.julia/packages/LoopVectorization/DDH6Z/src/reconstruct_loopset.jl:965
  [3] π’œπ’Έπ“‰!
    @ ~/.julia/packages/Tullio/NGyNM/src/macro.jl:1093 [inlined]
  [4] tile_halves(fun!::AttentionBasedPlmDCA.var"#π’œπ’Έπ“‰!#814", ::Type{Array{Float64}}, As::Tuple{OffsetArrays.OffsetArray{Float64, 3, Array{Float64, 3}}, Array{Float64, 3}, Matrix{Float64}, UnitRange{Int64}}, Is::Tuple{UnitRange{Int64}, UnitRange{Int64}, UnitRange{Int64}}, Js::Tuple{UnitRange{Int64}}, keep::Nothing, final::Bool)
    @ Tullio ~/.julia/packages/Tullio/NGyNM/src/threads.jl:139
  [5] tile_halves
    @ ~/.julia/packages/Tullio/NGyNM/src/threads.jl:136 [inlined]
  [6] threader
    @ ~/.julia/packages/Tullio/NGyNM/src/threads.jl:65 [inlined]
  [7] (::AttentionBasedPlmDCA.var"#ℳ𝒢𝓀ℯ#817"{AttentionBasedPlmDCA.var"#π’œπ’Έπ“‰!#814"})(wdelta::Array{Float64, 3}, probnew::Matrix{Float64}, β‰ͺ1:q≫::UnitRange{Int64})
    @ AttentionBasedPlmDCA ~/CODE/AttentionBasedPlmDCA.jl/src/wipattention.jl:140
  [8] (::Tullio.Eval{AttentionBasedPlmDCA.var"#ℳ𝒢𝓀ℯ#817"{AttentionBasedPlmDCA.var"#π’œπ’Έπ“‰!#814"}, AttentionBasedPlmDCA.var"#8942#βˆ‡β„³π’Άπ“€β„―#816"{AttentionBasedPlmDCA.var"#βˆ‡π’œπ’Έπ“‰!#815"}})(::Array{Float64, 3}, ::Vararg{Any})
    @ Tullio ~/.julia/packages/Tullio/NGyNM/src/eval.jl:20
  [9] mytest_tullio(wdelta::Array{Float64, 3}, probnew::Matrix{Float64})
    @ AttentionBasedPlmDCA ~/CODE/AttentionBasedPlmDCA.jl/src/wipattention.jl:303
 [10] top-level scope
    @ REPL[17]:1

On my mac it does not error but produces a strange

julia> mytest_tullio(a,b) # on mac os x
4Γ—4Γ—2 OffsetArray(::Array{Float64, 3}, 1:4, 1:4, 1:2) with eltype Float64 with indices 1:4Γ—1:4Γ—1:2:
[:, :, 1] =
 0.718021  0.718021  0.718021  0.718021
 0.747013  0.747013  0.747013  0.747013
 0.596645  0.596645  0.596645  0.596645
 0.852708  0.852708  0.852708  0.852708

[:, :, 2] =
 0.33938   0.33938   0.33938   0.33938
 0.93867   0.93867   0.93867   0.93867
 0.341837  0.341837  0.341837  0.341837
 0.26481   0.26481   0.26481   0.26481

Which is an offset array (???) rather than a 4x4x2 Array{Float64,3}.

This is just a minimal nonworking example; I could do it in another way. But I find it odd:

  1. The error on linux
  2. The strange offset array output on mac

Before bothering the Tullio devs, maybe some of you see some evident mistakes on my side.

Thanks

A

2

With Tullio alone, ideally it should notice that b in 1:q has a literal 1, and hence can be made OneTo(q). Failing that, it ought to notice that OffsetArrays is not loaded, and hence it can’t make an array with UnitRange axis, so it should insert something like first(1:b)==1 ? OneTo(b) : error("you really need OffsetArrays"), but it also seems to miss that check.

I think that’s a bug. The work-around is to define b_range = axes(wdelta,3) and write b in b_range.

julia> function mytest_tullio(wdelta,probnew)
           q = size(wdelta,3)
           @tullio mat[a,b,j] := wdelta[j,m,a] * probnew[m,a] (b in 1:q)  verbose=true
       end
...

julia> a = rand(2,3,4); b= rand(3,4); mytest_tullio(a,b)
β”Œ Info: left index ranges
β”‚   a = Base.OneTo(4)
β”‚   b = 1:4
β””   j = Base.OneTo(2)
β”Œ Info: reduction index ranges
β””   m = Base.OneTo(3)
ERROR: MethodError: no method matching similar(::Array{Float64, 3}, ::Type{Float64}, ::Tuple{Base.OneTo{Int64}, UnitRange{Int64}, Base.OneTo{Int64}})

The error with LoopVectorization is separate. It ought to be possible to reproduce this without Tullio by writing out the loops; calling the macro with verbose=2 will print everything including the generated loops.

1 Like
3

Thanks @mcabbott (I did not want to bother the devs … but here it goes :wink: )

Just to understand:

The fact that it errors on Linux (and not on Mac) is (possibly) due to LoopVectorization.

The fact it produces an OffsetArray is just a β€œcorner-case” not spotted by Tullio, right?

Anyway thanks a lot for your answer and most of all for the beautiful Tullio package!!!

A

1 Like