Hi,
I am having a strange problem with the tests of one of my packages, MessyTimeSeriesOptim.jl, which is currently under registration.
I have cloned the dev branch on my laptop and added the package via Pkg.add(path="..."). All tests pass if I run them via include("./test/runtests.jl"). However, if I use the more usual ]test MessyTimeSeriesOptim they fail at the testset "DFM simulations: MLE".
Why is ]test behaving differently from include("./test/runtests.jl")?
I have been once solving problem like this and it was caused by bounds checking. When you run the code with include, it inherits bounds checking from julia, if you test by ]test bounds are always checked, even though there is an @inbounds, i.e. it is overrided.
I see, thanks. However, I am not getting an error, it simply returns different floats at the end of the run. Is there anything else that gets overwritten with ]test?
]test runs with boundschecking forced to on (--check-bounds=yes).
If I include runtests.jl with --check-bounds=yes I get the same problem. However, if I remove all @inbounds and @turbo in my code and include runtests.jl without --check-bounds=yes all tests pass. Is --check-bounds=yes doing anything else? Can it affect one of my dependencies rather than my package directly? Also, is it possible to highlight which lines behave differently with -check-bounds=yes?
The ones with @inbounds. Usually the numeric difference comes from the code being able to use SIMD or not depending on the bounds checking behaviour.
Right, but I removed all @turbo and @inbounds now, and I am not using @simd directly. Can it be caused by some dependency?
It’s actually places that use @simd that are interesting. Could be in Base.
Thanks. I have double checked everything and the differences are minimal with/without check bounds. However, I was using a very small tolerance for the tests that were failing. I suppose that these small numerical differences are coming from somewhere in Base.