Is there something like broadcast_mapreduce?

General Usage
14

Thanks for the hint, I tried to adapt the expansion framework. I tried two approaches:

  1. An iterator similar to zip, but based on the broadcast expansion to repeat singleton dimensions
  2. A broadcast_reduce function, similar to mapreduce

Here are the results of some simple benchmarks:

using BenchmarkTools

x1 = rand(200,5000,30)
x2 = rand(1,5000,1)
x3 = rand(1,1,1)

# The allocating version
@btime sum(x1.*x2.*x3)
# 60.590 ms (57 allocations: 228.88 MiB)

# Iterator-based mapreduce
it = BroadIter(x1,x2,x3)
@btime mapreduce(x->x[1]*x[2]*x[3],+,it)
# 117.096 ms (3 allocations: 64 bytes)

# Iterator-based loop
function sum_iter(x1,x2,x3)
    it = BroadIter(x1,x2,x3)
    s = 0.0
    for (a1,a2,a3) in it
        s+=a1*a2*a3
    end
    s
end
@btime sum_iter(x1,x2,x3)
# 219.865 ms (4 allocations: 240 bytes)

# broadcast_reduce
@btime broadcast_reduce(*,+,0.0,x1,x2,x3)
# 134.958 ms (79 allocations: 1.92 KiB)

So the code is allocation-free and working correctly fro my use case, but still slower than the allocating version. I would guess that this is due to simd-optimizsations? I don’t think they are possible for the iterator approach. Any hints on how to improve one of the approaches would be welcome. The code that I adapted from broadcast.jl is here:

https://gist.github.com/meggart/3539a06f65f82d5149efcf90c37276b7