Michael Abbott

This is a fair point. And there's something to be said for matching `@time` (although that now adds compilation time, too). I'd also like to cut down `@benchmark` a bit,...

Isn't the mean what's relevant for throughput? I guess this PR takes the position that it's the most useful *second* number. For example, with the allocating `sum(x .* y)` above,...

Nice, somehow I never looked at the paper. I guess garbage collection is the major thing I had in mind, since it's very common to have different implementations with different...

> more detailed statistics, `@benchmark` is what you want But which bit -- what would you take from that as a better (if still imperfect) measure of the cost of...

So it looks like the blocks for the histogram aren't printing the same width as the ordinary space. Could try printing blocks set to the same colour as the background...

Ideally it would just call `float(p0)` on input, so that `curve_fit(model,x,y, [2,0])` assumes you meant `[2, 0.0]`. This will do nothing at all when they are already floats.

Xref https://github.com/JuliaDiff/ChainRulesCore.jl/issues/504

I wrote some tidier comments here: https://github.com/JuliaDiff/ChainRules.jl/pull/337#issuecomment-832080195 . This takes `norm` as the particular example, because I think it's helpful to be concrete, and times lots of things, because factors...

I think you miss some brackets: ``` julia> Zygote.gradient((d,m) -> sum(d*m), d, m)[1] 5×5 Matrix{Float64}: 3.24163 1.55452 2.24316 2.41607 2.82723 3.24163 1.55452 2.24316 2.41607 2.82723 3.24163 1.55452 2.24316 2.41607 2.82723...

Ok. So I agree you can imagine some automated rule that says "skip the abstract rule if there is a more specific primal". But I think there are two problems...