added reverse for arbitrary dimension CuArrays

[kraftpunk97-zz]

julia> test_shapes = [
         [1, 2, 4, 3],
         [4, 2],
         [5],
         [2^5, 2^5, 2^5],
       ]
4-element Array{Array{Int64,1},1}:
 [1, 2, 4, 3]
 [4, 2]      
 [5]         
 [32, 32, 32]

julia> for testshape ∈ test_shapes
           a = Float32.(reshape(Vector(1:prod(testshape)), testshape...))
           a_ = cu(a)
           b = similar(a)
           b_ = similar(a_)

           println("testing for shape: $testshape")
           for i=1:length(testshape)
               b = reverse(a, dims=i)
               my_reverse(a_, b_, i)
               @test all(cu(b) .== b_)
               println("dimension $i is ok.")
           end
           
           println()
       end
testing for shape: [1, 2, 4, 3]
dimension 1 is ok.
dimension 2 is ok.
dimension 3 is ok.
dimension 4 is ok.

testing for shape: [4, 2]
dimension 1 is ok.
dimension 2 is ok.

testing for shape: [5]
dimension 1 is ok.

testing for shape: [32, 32, 32]
dimension 1 is ok.
dimension 2 is ok.
dimension 3 is ok.

[kraftpunk97-zz]

benchmarks...

shape_ = [2^10, 2^10];
N = 2^10
a = Float32.(reshape(Vector(1:prod(shape_)), shape_...))# |> cu
a_ = cu(a)
b = similar(a)
b_ = similar(a_)

julia> @btime b =reverse(a, dims=2)
  740.175 μs (6 allocations: 4.00 MiB)

julia> @btime my_reverse(a_, b_, 2)
  129.330 μs (51 allocations: 1.83 KiB)

[kraftpunk97-zz]

pinging @vchuravy

[kraftpunk97-zz]

pinging @maleadt. Had this pull request going stale. Please review and suggest further changes, if any. Would be awesome if this could get merged 🙂

[maleadt]

Sorry for not responding. Pushed some small improvements.

@vchuravy we do need shared memory in the case input and output point to the same array (that is what we do with the other reverse kernel). That should probably be more clear by having two versions, one in-place with shared memory, and another out-of-place without.

[maleadt]

@kraftpunk97 could you check "allow pushes from maintainers"?

[kraftpunk97-zz]

It's already selected.

[maleadt]

Ugh, seems like I broke the PR. I'll open a new one...