Float32 print differently in matrices and vectors

[KristofferC]

julia> Float32[1.4013f-10 1]
1×2 Array{Float32,2}:
 1.4013e-10  1.0

julia> Float32[1.4013f-10, 1]
2-element Array{Float32,1}:
 1.4013f-10
 1.0f0  

1.4013e-10 vs 1.4013f-10.

[KristofferC]

Perhaps not a bug, but it looks a bit weird. Probably another consequence of #22981.

[rfourquet]

Yes this a direct consequence of it, and I already started to work on a fix. Note this is not at all limited to Float32, e.g.

julia> ModInt[ModInt{2}(0), ModInt{3}(0)]
2-element Array{Main.ModInts.ModInt,1}:
 0 mod 2
 0 mod 3

julia> ModInt[ModInt{2}(0) ModInt{3}(0)] 
1×2 Array{Main.ModInts.ModInt,2}:
 0  0

[rfourquet]

This example was also to show that currently in some cases the matrix printing looks better than vector's (Float32), but in other cases it seems wrong (ModInt). Also consider:

julia>  Number[Float16(1), Float32(1)]
2-element Array{Number,1}:
 Float16(1.0)
         1.0f0

julia> Number[Float16(1) Float32(1)]
1×2 Array{Number,2}:
 1.0  1.0

What would people think is better? The latter looses information, but maybe is sufficient?

[nalimilan]

It seems that this example complexifies the description of the problem proposed by @StefanKarpinski at #22981 (comment), since it adds a third dimension to take into account: consistency in printing, i.e. comparability of entries.

Indeed what's most annoying in the above example is that Float16(1.0) and 1.0f0 are printed in totally inconsistent styles, making them hard to compare. It's certainly fine to print an Array{Any} as a list of visually incomparable objects since they are supposed to be heterogeneous. But with Array{Number} you're likely to compare values with one another.

OTOH, if we dropped type information, this would be the only case where you wouldn't know the type of an entry in the context of array printing. It could be confusing for people who have not realized they got a heterogeneous array.

We could decide we don't care making this pretty, because arrays with abstract element types different from Any are quite rare, and use the explicit but ugly printing (as for vectors above). This would be less of an issue if we had a way to print 1.0f0 as Float32(1.0) for consistency with Float16(1.0).

[quinnj]

Looks like this is already solved:

julia> Float32[1.4013f-10, 1]
2-element Array{Float32,1}:
 1.4013e-10
 1.0

julia> Float32[1.4013f-10 1]
1×2 Array{Float32,2}:
 1.4013e-10  1.0