Milan Klöwer

Great, yeah, master now contains your version, as mine contained a bug. However, I get considerably slower timings ```julia julia> A = rand(Float32,300,300) .+ 1f0; # no subnormals julia> @btime...

I'd be happy if you want to add it, and I'm happy to assist implementing it. In principle there's only 2 things needed - Float8_5 to Float32 conversion (that's easy,...

From my thesis $n_e$, number of exponent bits, $n_m$, number of mantissa bits  with $bias = 2^{n_e-1} - 1$ and the fraction $$f = \sum_{i=1}^{n_m}m_i2^{-i}$$ (big endian). Then ```julia...

Given that we may want 3 different Float8 formats in this package, it might be worth not hardcoding the tables as we did before but to create them dynamically. It's...

Sorry, maybe to be a bit clearer, what I thought was to do something like ```julia abstract type AbstractFloat8

I was wondering that too. Maybe there are some tests that are skipped? Seems I haven’t been good at documenting this. Sorry. But I doubt it has been resolved.

So there are only 8 subnormals in Float8_4, 0000_0001 to 0000_0111 ```julia julia> for i in 0b0000_0001:0b0000_0111 f = Float8_4(i) println(bitstring(f)," ",f) end 00000001 Float8_4(0.001953125) 00000010 Float8_4(0.00390625) 00000011 Float8_4(0.005859375) 00000100...

So question is, where do these additional allocations come from? Could one not write a `nomissing` function similar to ```julia function ignoremissing(dsvar) # there's probably a more elegant way to...

Just realised that `in` cannot check for NaN's (see also https://github.com/JuliaLang/julia/issues/37157) in which case we probably need something like ```julia function value_in(val, collection) return !isnothing(findfirst(x -> x === val, collection))...

Yes SpeedyWeather uses bitrounding and netcdfs default zlib compression (deflate). It doesn't actually analyse the bitinformation it uses some default keepbits per variable.