Help is very welcome for any of the below:

• Automated optimisation / tuning of e.g. block_size for a given input; can be made algorithm-agnostic.

  • Maybe some thing like AK.@tune reduce(f, src, init=init, block_size=$block_size) block_size=(64, 128, 256, 512, 1024). Macro wizards help!

  • Or make it general like:

  AK.@tune begin
      reduce(f, src, init=init,
             block_size=$block_size,
             switch_below=$switch_below)
      block_size=(64, 128, 256, 512, 1024)
      switch_below=(1, 10, 100, 1000, 10000)
  end

• We need multithreaded implementations of sort, N-dimensional mapreduce (in OhMyThreads.tmapreduce) and accumulate (again, probably in OhMyThreads).

• Any way to expose the warp-size from the backends? Would be useful in reductions.

• Add a performance regressions runner.

• Other ideas? Post an issue, or open a discussion on the Julia Discourse.