Regression of reduce (formerly reducedim) in Julia 0.7

[wsshin]

I observe a significant regression in reduce in Julia 0.7 compared to reducedim in Julia 0.6. Below, I add the columns of an m×2 matrix to create an m×1 matrix for different values of m. In Julia 0.6,

julia> VERSION
v"0.6.3-pre.0"

julia> m = 1; A = @SMatrix rand(m, 2); @btime reducedim(+, $A, Val{2});
  1.894 ns (0 allocations: 0 bytes)

julia> m = 2; A = @SMatrix rand(m, 2); @btime reducedim(+, $A, Val{2});
  2.261 ns (0 allocations: 0 bytes)

julia> m = 3; A = @SMatrix rand(m, 2); @btime reducedim(+, $A, Val{2});
  2.261 ns (0 allocations: 0 bytes)

julia> m = 4; A = @SMatrix rand(m, 2); @btime reducedim(+, $A, Val{2});
  2.264 ns (0 allocations: 0 bytes)

julia> m = 5; A = @SMatrix rand(m, 2); @btime reducedim(+, $A, Val{2});
  2.634 ns (0 allocations: 0 bytes)

julia> m = 10; A = @SMatrix rand(m, 2); @btime reducedim(+, $A, Val{2});
  4.208 ns (0 allocations: 0 bytes)

On the other hand, in Julia 0.7:

julia> VERSION
v"0.7.1-pre.0"

julia> m = 1; A = @SMatrix rand(m, 2); @btime reduce(+, $A, dims=Val(2));
  8.015 ns (0 allocations: 0 bytes)

julia> m = 2; A = @SMatrix rand(m, 2); @btime reduce(+, $A, dims=Val(2));
  8.205 ns (0 allocations: 0 bytes)

julia> m = 3; A = @SMatrix rand(m, 2); @btime reduce(+, $A, dims=Val(2));
  210.229 ns (3 allocations: 160 bytes)

julia> m = 4; A = @SMatrix rand(m, 2); @btime reduce(+, $A, dims=Val(2));
  55.512 ns (2 allocations: 128 bytes)

julia> m = 5; A = @SMatrix rand(m, 2); @btime reduce(+, $A, dims=Val(2));
  44.934 ns (2 allocations: 144 bytes)

julia> m = 10; A = @SMatrix rand(m, 2); @btime reduce(+, $A, dims=Val(2));
  57.598 ns (2 allocations: 272 bytes)

Observations:

• 0.6 does not use any allocations.
• 0.7 starts using allocations for $m ≥ 3$, but even for $m ≤ 2$ for which no allocations are used, 0.7 is 3–4 times slower than 0.6.
• There is something weird about $m = 3$ in 0.7: the code runs significantly slower for $m = 3$ than for $m > 3$, probably because it uses one more allocation for some reason.

Any idea why this regression occurs?

[wsshin]

Maybe related to #439?

[tkoolen]

See also #494.

[nlw0]

When I reproduced your tests today I only saw allocations for m>=6, but sooner in my use case (4x4 matrices)

julia> VERSION
v"1.2.0-DEV.66"

julia> A = randn(4,4);

julia> for m = 1:10
display(m)
A = @smatrix rand(m, 2); @Btime reduce(+, $A, dims=Val(1));
A = @smatrix rand(m, m); @Btime reduce(+, $A, dims=Val(1));
end
1
1.814 ns (0 allocations: 0 bytes)
1.693 ns (0 allocations: 0 bytes)
2
1.695 ns (0 allocations: 0 bytes)
2.024 ns (0 allocations: 0 bytes)
3
1.754 ns (0 allocations: 0 bytes)
2.029 ns (0 allocations: 0 bytes)
4
2.031 ns (0 allocations: 0 bytes)
36.787 ns (2 allocations: 192 bytes)
5
2.032 ns (0 allocations: 0 bytes)
41.070 ns (2 allocations: 256 bytes)
6
37.657 ns (2 allocations: 144 bytes)
53.786 ns (2 allocations: 368 bytes)
7
32.522 ns (2 allocations: 160 bytes)
65.344 ns (2 allocations: 464 bytes)
8
34.810 ns (2 allocations: 176 bytes)
78.683 ns (2 allocations: 624 bytes)
9
35.368 ns (2 allocations: 192 bytes)
92.158 ns (2 allocations: 752 bytes)
10
38.982 ns (2 allocations: 208 bytes)
107.778 ns (2 allocations: 912 bytes)

Are there any open plans to solve this? This looks like a problem I might go around by myself in my own project, but it would be certainly better to have this solved in the library. Is there some way a newbie could help?

[andyferris]

Yes, this is annoying. It is possible that the problem might relate to how keyword argument functions are lowered - I think there was an issue or comment somewhere to automatically add @propagate_inbounds to the helper functions to fix certain inlining and boundschecking performance issues.

PS - Personally, I don't love the dims keyword appearing everywhere in method signatures since it reflects poor seperation of concerns. I would we write things like sum.(splitdims(A, 1)) (or the new eachrow, eachcol, eachslice functions).

[tkoolen]

Completely agree, I wish the dims kwarg API would just go away.

[c42f]

Probable same root cause as #540

[mateuszbaran]

One possible workaround for this is directly using the StaticArrays._reduce method like this:

f_(A) = StaticArrays._reduce(+, A, Val(2), NamedTuple())

That method was added in #659.