Replace tdma with built-in Julia implementation

[charleskawczynski]

🎉

[charleskawczynski]

I locally ran

function tridiag_solve(nz, b_rhs, a, b, c)
    # Note that `1:end` is zero-based indexing.
    A = LinearAlgebra.Tridiagonal(a[1:end], parent(b), c[0:(end - 1)])
    x_new = off_arr(A \ parent(b_rhs))
    x = deepcopy(b_rhs)
    tridiag_solve_old(nz, x, a, b, c)
    if !(x ≈ x_new)
        error("Not matching")
        @show a
        @show b
        @show c
        @show b_rhs
        @show x
        @show x_new
    else
    end
    return x
end


function tridiag_solve_old(nz, x, a, b, c)
    scratch = pyzeros(length(x))
    scratch[0] = c[0] / b[0]
    x[0] = x[0] / b[0]
    @inbounds for i in xrange(1, nz)
        m = 1.0 / (b[i] - a[i] * scratch[i - 1])
        scratch[i] = c[i] * m
        x[i] = (x[i] - a[i] * x[i - 1]) * m
    end
    # TODO: verify translation
    @inbounds for i in revxrange(nz - 2, 0, -1)
        x[i] = x[i] - scratch[i] * x[i + 1]
    end
    return
end

and found that the error is never hit, so the mse table changes are required due to machine precision errors.

[haakon-e]

Looks good.

[charleskawczynski]

bors r+

[bors]

Build succeeded:

• buildkite/turbulenceconvection-ci
• ci 1.5 - macOS-latest
• ci 1.5 - ubuntu-latest
• ci 1.5 - windows-latest
• ci 1.6.0 - macOS-latest
• ci 1.6.0 - ubuntu-latest
• ci 1.6.0 - windows-latest
• docbuild
• format