Unexpected behaviour when summing with different sign about tullio.jl HOT 2 CLOSED

Yes I think this is the correct behaviour. Here's another way to see it, same expressions:

julia> arr = rand(100,200);

julia> @tullio x2[i1, i2] := (arr[2i1 + 1, 2i2 + 1] + arr[2i1 + 0, 2i2 + 1] + arr[2i1 + 1, 2i2 + 0] + arr[2i1 + 0, 2i2 + 0])  verbose=true grad=false threads=false;
┌ Info: left index ranges
│   i1 = 1:49
└   i2 = 1:99

julia> @tullio x2[i1, i2] := (arr[2i1 - 1, 2i2 - 1] + arr[2i1 + 0, 2i2 - 1] + arr[2i1 - 1, 2i2 + 0] + arr[2i1 + 0, 2i2 + 0])  verbose=true grad=false threads=false;
┌ Info: left index ranges
│   i1 = 1:50
└   i2 = 1:100

2i1 + const means the range of i1 has to be about half of size(arr,1), which it is. In fact the range is something like i1 in (axes(arr,1) .- const) ./ 2, but this division has to be done carefully. And then it takes the intersection over the different terms. In all of these cases, the previous/next integer not included in the ranges it chose would put you outside of axes(arr,1) in one of the terms.

from tullio.jl.

Thanks for your explanation, it seems very reasonable 🥰

I'll close that issue.

from tullio.jl.