Tests take too long

[tkoolen]

We're getting timeouts on Travis. Could work around the problem using travis_wait, but I think it would be better to just reduce the time it takes to run the tests.

[tkoolen]

Started investigating this a bit. I thought I'd just get rid of the N = 14 case in some of the new tests, but:

julia> using StaticArrays, Compat.Test

julia> @time @testset "Matrix inverse NxN, N > 14" begin
           for sz in (15, 50), typ in (Float64, Complex{Float64})
               A = rand(typ, sz, sz)
               SA = SMatrix{sz,sz,typ}(A)
               @test inv(A) ≈ inv(SA)
           end
       end
Test Summary:              | Pass  Total
Matrix inverse NxN, N > 14 |    4      4
182.003208 seconds (296.32 M allocations: 8.784 GiB, 0.85% gc time)
Base.Test.DefaultTestSet("Matrix inverse NxN, N > 14", Any[], 4, false)

182 seconds for cases where the generic fallback is called.

[andyferris]

There's always the technique of having test code of different lengths, and have CI only run faster tests. If you wanted to split out a bunch of tests into a longtests directory or something where we can do extensive fuzz testing, that might solve our problem.

[tkoolen]

It turns out that it's not really the inv that's slow to compile for large matrices; rather it's the isapprox call!

julia> using StaticArrays

julia> A = rand(Complex{Float64}, 50, 50);

julia> SA = SMatrix{50,50,Complex{Float64}}(A); # 8 seconds

julia> @time isapprox(A, SA)
185.520908 seconds (256.89 M allocations: 7.525 GiB, 0.83% gc time)

I'll figure out why and try to fix it.

Edit: vecnorm is the main culprit.

[tkoolen]

The bigger issue is just that lots of functions currently don't have cutoffs for when to fall back to Base methods. I think a quick fix where we just avoid testing such big matrices is appropriate for now.