ExBroadcast.jl

• Extend Base.Broadcast by macros:
  • @tab: Tuple of Array Broadcast --- broadcast with multiple outputs will be stored in tuple of array (instead of array of tuple).
  • @mtb: MultiThread Broadcast --- perform broadcast with multiple threads.
  • @mtab: @mtb + @tab
  • @stb: force STructArray Broadcast --- it only works if user loads StructArrays.jl

Macros' Usage

1. @tab : support CuArray, OffsetArray, Tuple, StructArray, StaticArray

julia> a = randn(4000,4000);
julia> @tab b, c = sincos.(a);
julia> @tab b, c = broadcast(sincos,a);
julia> @tab b, c = broadcast(a) do x
          sincos(x)
       end;
julia> @tab b, c .= sincos.(a);
julia> broadcast!(sincos,(b,c),a);

• For outputs <: AbstractArray
  • Only the default copy method which use similar(bc, T) is implemented, thus inputs like StaticArray is not allowed for non-inplace caluculation by default. We have an extension for @tab with StaticArrays.
  • @tab is not optimized for BitArray. The default return type is Array{Bool} for non-inplace broadcast.
• For outputs <: Tuple, @tab first generate all results and then seperate them.
• @tab is not designed for too many outputs.

2. @mtb : cpu multi-threads broadcast

julia> a = randn(4000,4000); b = similar(a);
julia> @btime @mtb @. $b = sin(a);
  47.756 ms (22 allocations: 2.97 KiB)
julia> @btime @. $b = sin(a);
  167.985 ms (2 allocations: 32 bytes)
julia> Threads.nthreads()
 4

• @mtb use CartesianPartition to seperate the task with dimension > 1
• @mtb will be turned off automately for CuArray and Tuple
• @mtb assume all elements in the dest array(s) are seperated in the memory and there's no thread safety check.
• @mtb is not tuned for small arrays (It won't invoke the single thread version automately).
• User can change the number of threads by :
  • Call ExBroadcast.set_num_threads(n) for global change.
  • Use 2 inputs macro @mtb n [...] for local change. (thread safe)

Note

1. @mtab only save some compile cost.