promote fillvalue type if it is Colorant

[johnnychen94]

It's weird that regions are not filled by fillvalue in following cases

# in master: filled by Gray(RGB(1.,0.,0.))
# in this PR: filled by RGB(1.,0.,0.)
PaddedView(RGB(1.,0.,0.), fill(Gray(0.), 2, 2), (-1:4, -1:4))

Cases such as transparent colors as fillvalue are also handled.

The only issue I see is the loading time:

• this branch without ColorTypes loaded: 0.26s
• this branch with ColorTypes loaded: 0.48s
• master: 0.07s

If the loading time is unacceptable, the only place I can think of is ImageShow/ImageCore -- but that would be type piracy IIUC.

This PR serves JuliaImages/ImageShow.jl#19 to make mosaicview a really useful imshow alternative

[johnnychen94]

I now think this patch should go to ImageCore instead of this package.

[johnnychen94]

closed in favor of #26

[timholy]

Sorry to be slow to look at this, "real work" has been quite insane lately. Next week should settle out considerably.

What about adding a constructor PaddedView{T}(fillvalue, data, ...) letting the user pick the eltype directly?

[johnnychen94]

Sorry to be slow to look at this, "real work" has been quite insane lately. Next week should settle out considerably.

Keep safe! ❤️

What about adding a constructor PaddedView{T}(fillvalue, data, ...) letting the user pick the eltype directly?

Exporting the type control to users looks good to me, but I'm a little worried when this usage being abused, e.g.,

julia> A = rand(Float64, 512, 512);

julia> @btime collect(PaddedView(0, $A, (-2:515, -2:515)));
  167.392 μs (3 allocations: 2.05 MiB)

julia> @btime collect(PaddedView{Float32}(0, $A, (-2:515, -2:515)));
  195.014 μs (3 allocations: 1.02 MiB)

would this be an issue?

[timholy]

In the first case values are being returned directly from A, in the second they are converted to Float32 before use. It's slightly puzzling in light of

julia> using BenchmarkTools, MappedArrays

julia> A = rand(Float64, 512, 512);

julia> B64 = of_eltype(Float64, A);

julia> B32 = of_eltype(Float32, A);

julia> @btime collect($B64);
  106.468 μs (2 allocations: 2.00 MiB)

julia> @btime collect($B32);
  71.884 μs (2 allocations: 1.00 MiB)

but the presence of a branch here is not to be discounted.

[johnnychen94]

I think I benchmarked incorrectly, the memory reallocation differs for Float32 and Float64. The new benchmark is left in #27