Make mapCube/xmap mathods specific for array type

[danlooo]

Currently, there is only one method of mapCube and xmap for a YAXArray.
Do we want to add xmap methods for each array type, e.g. Matrix, DiskArrays.CachedDiskArray so that the three calls above perform similar? Seems that mapCube treats everything like a DiskArray.

using YAXArrays
using BenchmarkTools

data = ones(1000, 1000)
a = YAXArray(data)

@benchmark map(x -> 5, a)

@benchmark mapCube(
    a,
    indims=InDims(:Dim_1, :Dim_2),
    outdims=OutDims(a.Dim_1, a.Dim_2),
) do xout, xin
    xout .= 5
end

@benchmark mapCube(
    cache(a),
    indims=InDims(:Dim_1, :Dim_2),
    outdims=OutDims(a.Dim_1, a.Dim_2),
) do xout, xin
    xout .= 5
end

Resulting in

julia> @benchmark map(x -> 5, a)
BenchmarkTools.Trial: 8181 samples with 1 evaluation per sample.
 Range (min … max):  441.142 μs …   3.996 ms  ┊ GC (min … max):  0.00% … 63.13%
 Time  (median):     520.456 μs               ┊ GC (median):     0.00%
 Time  (mean ± σ):   607.869 μs ± 255.064 μs  ┊ GC (mean ± σ):  14.72% ± 18.98%

  ▇▆▅▄██▆▅▄▃▁                           ▂▂▁▁ ▁▂▄▄▃▂▁            ▂
  ████████████▇▅▆▃▄▄▄▃▁▁▁▁▅▁▄▁▄▁▁▁▁▁▃▅▆██████████████▇▆▆▆▅▅▆▅▅▆ █
  441 μs        Histogram: log(frequency) by time        1.4 ms <

 Memory estimate: 7.63 MiB, allocs estimate: 29.

julia> @benchmark mapCube(
           a,
           indims=InDims(:Dim_1, :Dim_2),
           outdims=OutDims(a.Dim_1, a.Dim_2),
       ) do xout, xin
           xout .= 5
       end
BenchmarkTools.Trial: 76 samples with 1 evaluation per sample.
 Range (min … max):  61.399 ms … 103.611 ms  ┊ GC (min … max): 91.74% … 92.86%
 Time  (median):     65.496 ms               ┊ GC (median):    91.27%
 Time  (mean ± σ):   66.599 ms ±   6.308 ms  ┊ GC (mean ± σ):  91.48% ±  1.10%

  ▃  ▂█▂  ▂ ▃▃▂                                                 
  █▁▇███▇▇█████▇▅▄▇▁▁▄▁▁▁▄▁▄▁▁▁▁▁▄▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▄▁▄ ▁
  61.4 ms         Histogram: frequency by time         89.2 ms <

 Memory estimate: 32.46 MiB, allocs estimate: 932.

julia> @benchmark mapCube(
           cache(a),
           indims=InDims(:Dim_1, :Dim_2),
           outdims=OutDims(a.Dim_1, a.Dim_2),
       ) do xout, xin
           xout .= 5
       end
BenchmarkTools.Trial: 75 samples with 1 evaluation per sample.
 Range (min … max):  62.597 ms … 84.940 ms  ┊ GC (min … max): 88.99% … 88.09%
 Time  (median):     66.699 ms              ┊ GC (median):    88.10%
 Time  (mean ± σ):   67.298 ms ±  4.022 ms  ┊ GC (mean ± σ):  88.76% ±  1.84%

       ▁▆     ▁▃█ ▁                                            
  ▄▄▇▆▇██▇▇▆▇▇███▄█▄▄▆▇▆▄▄▁▄▁▁▁▄▁▁▁▁▄▁▁▁▁▁▁▁▁▁▁▁▄▁▁▁▁▁▁▁▁▁▁▄▄ ▁
  62.6 ms         Histogram: frequency by time        81.6 ms <

 Memory estimate: 40.09 MiB, allocs estimate: 997.

[meggart]

I am not really sure what we are benchmarking here. Why do we read the array a at all if all we want to do is create a new YAXArray with constant values? The compiler can do some shortcuts when it know that the function you apply does not depend on x at all, note that @benchmark map(x -> x, a) is slower than . Also we are currently comparing an implementation of the new (map, which would fall back to DiskArrayEngine if the wrapped array is a DiskArray) against old (mapCube) behavior.