Zirui He

Hi Chris and ghost, there seems another issue when a function is used as the coeff_fun. I tried the example in UpwindDifference: > drift = [1., 1., -1.] > UpwindDifference(1,...

thanks for fixing the issue. However, I found an issue concerning UpwindDifference when the coeff_fun is not given. For example: `julia> drift=[1.0, 1., -1., 1., -1.]` `julia> BandedMatrix(UpwindDifference(1,1,0.1,5,drift))` gives correct...

I confirm this problem is still there. And it's very strange that it takes such long time for such small, 5 * 5, matrix. And surprisingly, when you do this...

> It may be a threading performance issue. We may need to use `@turbo` or whatnot. Hi Chris, did you notice the difference in allocations between operating the whole matrix...

this method ``` function log_nan(x::T)::T where {T @benchmark NaNMath.log(-0.1) BenchmarkTools.Trial: 10000 samples with 1000 evaluations. Range (min … max): 2.800 ns … 123.000 ns ┊ GC (min … max): 0.00%...

@MilesCranmer Hi Miles, it is about the method from @johanbluecreek