Fix accumulate bug (#2005)

Author: chrstphrbrns

src/accumulate.jl

@@ -177,7 +177,9 @@ function scan!(f::Function, output::AnyCuArray{T}, input::AnyCuArray;
 
         # get the total of each thread block (except the first) of the partial scans
         aggregates = fill(neutral, Base.setindex(size(input), blocks_dim, dims))
-        copyto!(aggregates, selectdim(output, dims, partial:partial:length(Rdim)))
+        partials = selectdim(output, dims, partial:partial:length(Rdim))
+        indices = CartesianIndices(partials)
+        copyto!(aggregates, indices, partials, indices)
 
         # scan these totals to get totals for the entire partial scan
         accumulate!(f, aggregates, aggregates; dims=dims)

test/base/array.jl

@@ -318,6 +318,7 @@ end
 @testset "accumulate" begin
   for n in (0, 1, 2, 3, 10, 10_000, 16384, 16384+1) # small, large, odd & even, pow2 and not
     @test testf(x->accumulate(+, x), rand(n))
+    @test testf(x->accumulate(+, x), rand(n,2))
     @test testf((x,y)->accumulate(+, x; init=y), rand(n), rand())
   end