Michael Abbott

Looks like https://github.com/JuliaLang/julia/issues/29393 is the issue about dangling type parameters. Will fiddle a bit more first, but PaddedMatricesForwardDiff doesn't sound like a bad option to store these methods, since it...

Perhaps of interest: ``` using Tullio bar(x) = sum(@tullio y[i,j] := tanh(x[i,j])) bar'(x) bar2(x) = @tullio s := tanh(x[i,j]) # without allocating y bar2'(x) ```

Thanks, yes I can see how all orders of 8 loops is quite a few to check! There is indeed low-hanging fruit, in that splitting this up into successive multiplications...

Right, the wrong answer is my fault, sorry I should have said. But I don't see why yet. That unrolling helps a lot. If I'm counting right, these nested loops...

That's pretty qiuck. LoopVectorization is doing some unnrolling then, but not as much? Good catch on the counting, I missed a factor 4. BTW, am I right to think that...

These things always seem noisy. It does say "4 of 4 new or added lines in 1 file covered. (100.0%)"

After rounding, shouldn't the test be approximate equality at the resolution you chose? The default is about 8 digits, and you've declared that you don't care about most of them....

Maybe worth mentioning here that, at least for the simplest case in which you have dense arrays of homogeneous units, you can just reinterpret them: ```julia function LinearAlgebra.:\(A::Matrix{Q}, b::Vector{Q}) where...

I fixed a few, not sure I understand the package well enough to fix all of them, nor to know whether they can ever matter in practice. So maybe this...

Sure. The more refined version of this is https://github.com/PainterQubits/Unitful.jl/pull/328#issuecomment-924382532 : After declaring that you care about 3 digits, how surprising is it really that things don't agree beyond that? ("3...