Very slow 4D broadcast in 2.0.1

[DrChainsaw]

Describe the bug
I get very poor performance on broadcasting operations on arrays with more than 3 dimensions compared to 1.7.3

To Reproduce
The Minimal Working Example (MWE) for this bug:

## 1.7.3
julia> using CuArrays

julia> using BenchmarkTools

julia> CuArrays.allowscalar(false)

julia> const xxxx = cu(randn(64,64,64,64));

julia> const yyyy = cu(randn(1,1,1,64));

julia> @benchmark CuArrays.@sync xxxx .+ yyyy
BenchmarkTools.Trial: 
  memory estimate:  4.20 KiB
  allocs estimate:  78
  --------------
  minimum time:     504.800 μs (0.00% GC)
  median time:      584.401 μs (0.00% GC)
  mean time:        718.303 μs (1.35% GC)
  maximum time:     4.462 ms (37.51% GC)
  --------------
  samples:          6871
  evals/sample:     1

(CuArraysFp16) pkg> up CuArrays
  Updating registry at `E:/Programs/julia/.julia/registries/General`
  Updating git-repo `https://github.com/JuliaRegistries/General.git`
 Resolving package versions...
  Updating `E:\Programs\julia\.julia\dev\CuArraysFp16\Project.toml`
  [3a865a2d] ↑ CuArrays v1.7.3 ⇒ v2.0.1
  Updating `E:\Programs\julia\.julia\dev\CuArraysFp16\Manifest.toml`
  [3895d2a7] ↑ CUDAapi v3.1.0 ⇒ v4.0.0
  [c5f51814] ↑ CUDAdrv v6.0.0 ⇒ v6.2.2
  [be33ccc6] ↑ CUDAnative v2.10.2 ⇒ v3.0.3
  [da1fd8a2] + CodeTracking v0.5.8        
  [f68482b8] + Cthulhu v1.0.1
  [3a865a2d] ↑ CuArrays v1.7.3 ⇒ v2.0.1
  [e2ba6199] + ExprTools v0.1.0        
  [0c68f7d7] ↑ GPUArrays v2.0.1 ⇒ v3.1.0
  [189a3867] + Reexport v0.2.0
  [76f85450] + LibGit2 
  [44cfe95a] + Pkg 
  [3fa0cd96] + REPL 

julia> using CuArrays

julia> using BenchmarkTools

julia> CuArrays.allowscalar(false)

julia> const xxxx = cu(randn(64,64,64,64));

julia> const yyyy = cu(randn(1,1,1,64));

julia> @benchmark CuArrays.@sync xxxx .+ yyyy
BenchmarkTools.Trial: 
  memory estimate:  4.30 KiB
  allocs estimate:  87
  --------------
  minimum time:     20.733 ms (0.00% GC)
  median time:      25.089 ms (0.00% GC)
  mean time:        24.637 ms (0.02% GC)
  maximum time:     29.447 ms (0.00% GC)
  --------------
  samples:          203
  evals/sample:     1

Expected behavior
I guess I was hoping for performance to at least be on par with 1.7.3 :)

Build log

julia> Pkg.build()
  Building NNlib → `E:\Programs\julia\.julia\packages\NNlib\FAI3o\deps\build.log`
false

julia> using CuArrays, BenchmarkTools
[ Info: Precompiling CuArrays [3a865a2d-5b23-5a0f-bc46-62713ec82fae]

julia> const xxxx = cu(randn(64,64,64,64));

julia> const yyyy = cu(randn(1,1,1,64));

julia> @benchmark CuArrays.@sync xxxx .+ yyyy
BenchmarkTools.Trial: 
  memory estimate:  4.30 KiB
  allocs estimate:  87
  --------------
  minimum time:     19.186 ms (0.00% GC)
  median time:      25.103 ms (0.00% GC)
  mean time:        24.327 ms (0.03% GC)
  maximum time:     28.657 ms (5.64% GC)
  --------------
  samples:          206
  evals/sample:     1

Environment details (please complete this section)
Details on Julia:

julia> versioninfo()
Julia Version 1.3.1
Commit 2d5741174c (2019-12-30 21:36 UTC)       
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Core(TM) i7-5820K CPU @ 3.30GHz
  WORD_SIZE: 64    
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, haswell)

Julia packages:
See MWE above

CUDA: toolkit and driver version: Driver version 442.74. CUDA bundled with 2.0.1? I have 10.2 installed for 1.7.3.

Additional context
Add any other context about the problem here.

[maleadt]

Confirmed, bisected to JuliaGPU/CUDAnative.jl@cf8f55b