PERF: BSplineKernelFunction::FastEvaluate(u): a faster way to evaluate BSplineKernelFunction

[N-Dekker]

Added a static member function BSplineKernelFunction::FastEvaluate(u). Equivalent to the existing BSplineKernelFunction::Evaluate(u) member function. Allows writing faster code, because it is non-virtual, and
because it does not require the creation of a BSplineKernelFunction object.

Removed the dynamically allocated m_Kernel data member from BSplineInterpolationWeightFunction and replaced the virtual function call m_Kernel->Evaluate(x) by a faster static member function call, to the newly added BSplineKernelFunction::FastEvaluate member function.

[dzenanz]

Can you comment on the performance improvement @N-Dekker?

[N-Dekker]

Can you comment on the performance improvement @N-Dekker?

@dzenanz Not yet, but I'm pretty sure that removing both the dynamic allocation of the kernel and the virtual function call does improve the performance.

Update: I just compared old style kernel->Evaluate...

    double sum{};
    const auto kernel = BSplineKernelFunctionType::New();

    for (int i{}; i < n; ++i)
    {
      sum += kernel->Evaluate(sum);
    }

...to calling the proposed static member function, FastEvaluate:

    double sum{};

    for (int i{}; i < n; ++i)
    {
      sum += BSplineKernelFunctionType::FastEvaluate(sum);
    }

Using VS2019 x64 (Release). For SplineOrder = 3, the duration was reduced by more than 8% when calling FastEvaluate. For SplineOrder = 1, the duration was even reduced by 55%!