@Turbo failing

[ParadaCarleton]

For some reason, @turbo throws a warning here at this specific line in my code. (Note that this was failing before I made a PR that set avx=false to silence the warning for users.) I have no idea what's going on. I haven't been able to build a MWE because from what I can see the line is so simple that I feel like it should definitely be running; it's probably to do with some weird edge case that I can't isolate without knowing the internals of @tullio.

[mcabbott]

I'm surprised that fails, it's so simple. But maybe something is tripping over the type of to_sum? Can you post the error message, or the warning?

[ParadaCarleton]

I'm surprised that fails, it's so simple. But maybe something is tripping over the type of to_sum? Can you post the error message, or the warning?

Bizarrely, it's disappeared now. I have no idea what changed and fixed this, but I can't reproduce it anymore.

[ParadaCarleton]

I'm surprised that fails, it's so simple. But maybe something is tripping over the type of to_sum? Can you post the error message, or the warning?

Oh weird, it's popping up again. Maybe I just missed it? Here's the warning:

┌ Warning: #= /home/lime/.julia/packages/Tullio/qPZkO/src/macro.jl:1094 =#:
│ `LoopVectorization.check_args` on your inputs failed; running fallback `@inbounds @fastmath` loop instead.
│ Use `warn_check_args=false`, e.g. `@turbo warn_check_args=false ...`, to disable this warning.
└ @ ParetoSmooth ~/.julia/packages/LoopVectorization/atM0u/src/condense_loopset.jl:822

[chriselrod]

The warning will only print once by default. You can make it appear more often via @turbo warn_check_args=5, which will let it print 5 times (accepts any value that can be converted to an Int).

[mcabbott]

Ideally Tullio's own check should not send LV things it can't handle. But perhaps it can be made smarter. What are the types of the arrays here?

[chriselrod]

What are the types of the arrays here?

I should probably have LV's warnings print those.

[ParadaCarleton]

Ideally Tullio's own check should not send LV things it can't handle. But perhaps it can be made smarter. What are the types of the arrays here?

The arrays are just normal Base arrays of Float64s, with an AxisKeys.jl KeyedArray wrapper around them.

[chriselrod]

with an AxisKeys.jl KeyedArray wrapper around them.

Okay, that's why.

julia> using LoopVectorization, AxisKeys

julia> A = KeyedArray(rand(2,10), ([:a, :b], 10:10:100))
2-dimensional KeyedArray(...) with keys:
↓   2-element Vector{Symbol}
→   10-element StepRange{Int64,...}
And data, 2×10 Matrix{Float64}:
        (10)          (20)         (30)         (40)          (50)         (60)          (70)         (80)         (90)         (100)
  (:a)     0.498766      0.363857     0.313591     0.303895      0.453926     0.0574657     0.443101     0.179118     0.39422       0.946817
  (:b)     0.0080239     0.544323     0.983931     0.0620305     0.589515     0.903863      0.114999     0.816759     0.377231      0.306352

julia> LoopVectorization.check_args(A)
false

For an array type to be supported, it must implement the ArrayInterface.jl stride array interface (or ArrayInterface can define the appropriate methods itself).

If LoopVectorization.check_args(A) == true and LoopVectorization.stridedpointer(A) returns a sensible result, then it should work. Currently the former returns false and the latter throws an error.

[ParadaCarleton]

with an AxisKeys.jl KeyedArray wrapper around them.

Okay, that's why.

julia> using LoopVectorization, AxisKeys

julia> A = KeyedArray(rand(2,10), ([:a, :b], 10:10:100))
2-dimensional KeyedArray(...) with keys:
↓   2-element Vector{Symbol}
→   10-element StepRange{Int64,...}
And data, 2×10 Matrix{Float64}:
        (10)          (20)         (30)         (40)          (50)         (60)          (70)         (80)         (90)         (100)
  (:a)     0.498766      0.363857     0.313591     0.303895      0.453926     0.0574657     0.443101     0.179118     0.39422       0.946817
  (:b)     0.0080239     0.544323     0.983931     0.0620305     0.589515     0.903863      0.114999     0.816759     0.377231      0.306352

julia> LoopVectorization.check_args(A)
false

Oh, hmm, that's surprising. It looks like it doesn't work with DimensionalData either. @rafaqz

[rafaqz]

@chriselrod what ArrayInterface.jl methods do we have to implement for this to work?

[chriselrod]

If it's just a simple wrapper, then all that is needed is

ArrayInterface.parent_type(::Type{<:YourArrayWrapper}) = ...

For example:

julia> using AxisKeys, ArrayInterface, LoopVectorization

julia> A = KeyedArray(rand(2,10), ([:a, :b], 10:10:100))
2-dimensional KeyedArray(...) with keys:
↓   2-element Vector{Symbol}
→   10-element StepRange{Int64,...}
And data, 2×10 Matrix{Float64}:
        (10)          (20)         (30)         (40)         (50)         (60)         (70)          (80)         (90)         (100)
  (:a)     0.0677124     0.755331     0.558675     0.213256     0.719723     0.195234     0.0587435     0.866621     0.42191       0.156516
  (:b)     0.761694      0.135312     0.386402     0.54806      0.994232     0.730863     0.128959      0.898721     0.971936      0.971892

julia> ArrayInterface.parent_type(::Type{<:KeyedArray{T,N,A}}) where {T,N,A} = A

julia> LoopVectorization.check_args(A)
true

julia> stridedpointer(A)
LayoutPointers.StridedPointer{Float64, 2, 1, 0, (1, 2), Tuple{Static.StaticInt{8}, Int64}, Tuple{Static.StaticInt{1}, Static.StaticInt{1}}}(Ptr{Float64} @0x00007f1f3a82f5a0, ArrayInterface.StrideIndex{2, (1, 2), 1, Tuple{Static.StaticInt{8}, Int64}, Tuple{Static.StaticInt{1}, Static.StaticInt{1}}}((static(8), 16), (static(1), static(1))))

julia> stridedpointer(A) === stridedpointer(parent(A))
true