cat can grow ndims. Fixes JuliaGPU#72, JuliaGPU#73

Unnecessary copies are created when growing ndims.

src/utils.jl

@@ -85,13 +85,22 @@ end
   x, ntuple(n -> n == dim ? i : I[n], Val{N})
 end
 
+function growdims(dim, x)
+  if ndims(x) >= dim
+    x
+  else
+    x[fill(Colon(), dim)...]
+  end
+end
+
 function _cat(dim, dest, xs...)
   function kernel(dim, dest, xs)
     I = @cuindex dest
     @inbounds n, I′ = catindex(dim, Int.(I), size.(xs))
     @inbounds dest[I...] = xs[n][I′...]
     return
   end
+  xs = growdims.(dim, xs)
   blk, thr = cudims(dest)
   @cuda (blk, thr) kernel(dim, dest, xs)
   return dest

test/runtests.jl

@@ -75,6 +75,8 @@ end
   @test testf(vcat, ones(5), zeros(5))
   @test testf(hcat, rand(3, 3), rand(3, 3))
   @test testf(vcat, rand(3, 3), rand(3, 3))
+  @test testf(hcat, rand(3), rand(3))
+  @test testf(cat, 4, rand(3, 4), rand(3, 4))
 end
 
 @testset "Broadcast" begin