support multiple return values?

[ssfrr]

I'm trying to slice-broadcast a function that has multiple return values. Ideally I'd collect them into separate arrays. Is this possible witih TensorCast?

Here's a MWE of basically what I'm trying to do:

x = randn(10, 20)
@cast v[k], i[k] := findmax(x[:, k]) 

But this throws the error:

LoadError: don't know what to do with left = (v[k], i[k])

Which is a great error message that leads me to believe I can't do this kind of thing, but I wanted to check.

In this case it's probably not too bad to just collect into a single Vector{Tuple{Float64, Int} and split it afterwards, but my actual example has some more complicated indexing stuff going on that would benefit from some TensorCast magic.

[ssfrr]

For a more concrete but less MWE example, here's what I'm doing right now:

@cast rank1res[k] := rank1filter(permutedims(specs[k, :, :]))
@cast sproj[k, n, c] := rank1res[k][1][c, n]
@cast vs[k, c] := rank1res[k][2][c]
@cast λs[k] := rank1res[k][3]
@cast q[k] := rank1res[k][4]

Though actually I'm not sure how I would get rid of the permutedims stuff here if it was combined with multiple return values, so maybe this isn't a great idea after all.

[mcabbott]

Have never thought about that.

Fundamentally this package wants to make one big array from the RHS, and then there's another set of procedures to reshape/slice as desired by the LHS, afterwards. What certainly ought to work (but doesn't) is this:

@cast vi[j,k] := findmax(x[:, k])[j] 

And what sort-of works is the following awful hack, because the LHS parser never checks that (v, i) isn't a symbol, and so just assigns to it like (v, i) = eachcol(rhs):

x = randn(10, 20);
Base.ndims(::Tuple) = 1
@cast (v, i)[j][k] := findmax(x[:, k])[j] lazy
v

[mcabbott]

FWIW, behaviour on version 0.4 is that these work without tweaks, and should probably be added to the tests:

julia> x = rand(Int8, 3, 5)
3×5 Matrix{Int8}:
  76  -128   57   -8  -52
 -85    15  -33  -55  -20
  -2    -2  -29  -23  -17

julia> @cast vi[j,k] := findmax(x[:, k])[j] 
2×5 Matrix{Signed}:
 76  15  57  -8  -17
  1   2   1   1    3

julia> @cast (v, i)[j][k] := findmax(x[:, k])[j]
2-element Vector{SubArray{Signed, 1, LinearAlgebra.Transpose{Signed, Matrix{Signed}}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64}, false}}:
 [76, 15, 57, -8, -17]
 [1, 2, 1, 1, 3]

julia> v
5-element view(transpose(::Matrix{Signed}), :, 1) with eltype Signed:
  76
  15
  57
  -8
 -17

julia> @cast (v, i)[j][k] |= findmax(x[:, k])[j]  # less lazy
2-element Vector{Vector{Signed}}:
 [76, 15, 57, -8, -17]
 [1, 2, 1, 1, 3]

julia> @pretty @cast (v, i)[j][k] := findmax(x[:, k])[j]
begin
    ndims(x) == 2 || throw(ArgumentError("expected a 2-tensor x[:, k]"))
    local sandpiper = sliceview(x, (:, *))
    local butterfly = @__dot__(findmax(sandpiper))
    local raven = transmute(stack(butterfly), (2, 1))
    (v, i) = sliceview(raven, (:, *))
end

Still not ideal in terms of the element types, I guess. With Float64:

julia> x = rand(3, 5);

julia> @cast vi[j,k] := findmax(x[:, k])[j] 
2×5 Matrix{Real}:
 0.853762  0.394276  0.739257  0.347495  0.804691
 3         2         3         3         1

julia> @cast (v, i)[j][k] := findmax(x[:, k])[j];

julia> Int.(i)
5-element Vector{Int64}:
 3
 2
 3
 3
 1

[mcabbott]

On this topic, StructArrays allows for efficient broadcasting out to several arrays, as for instance here: https://github.com/JuliaDiff/ChainRules.jl/blob/main/src/unzipped.jl#L6

It might not be hard to allow the pattern @cast v[k], i[k] := stuff[k] (no reduction, same indices on all both LHS arrays), implemented using something like this, or exactly with unzip_broadcast.

Issue about making this work with GPU arrays: JuliaArrays/StructArrays.jl#150