Work around GPU launch bug for BatchNorm

[staticfloat]

X-ref: #385

[MikeInnes]

Julia's escape analysis can make for some fun gpu debugging.

Do we have a test that catches this?

[staticfloat]

Yes; here you go:

using CuArrays, Flux

# Bogus initializations
μ = gpu(randn(Float32, 1, 1, 8, 1))
σ = μ

# It's important that these are tracked; it doesn't fail otherwise
β = param(μ)
γ = param(μ)

# Random input
x = gpu(randn(Float32, 10, 10, 8, 1))

# If any part of this computation is taken out or refactored, it does not error.
γ .* ((x .- μ) ./ σ) .+ β

[MikeInnes]

Actually, I guess this is not escape analysis but some failure of broadcast compilation when it gets this complex.

That's useful but also, does this issue get caught by the flux gpu tests now? If not we should just add something that calls batchnorm.

[KristofferC]

This content of this PR is in #586, can be closed.

[staticfloat]

Great, thanks @KristofferC!