PreconditionRelaxation: step w. relaxation

[peterrum]

Add specialization for preconditioners with:

void vmult(dst, src, relax);
void step(dst, src, relax);

[bangerth]

Looks good. It's time we get to use C++20 concepts :-)

I assume you're going to test this in a follow-up patch somewhere?

[bangerth]

For consistency with other place:

Suggested change

	std::declval<U const>().step(std::declval<VectorType &>(),
	std::declval<const U>().step(std::declval<VectorType &>(),

[bangerth]

Let's turn this around to make it read like "if this complicated type is 'bool'":

Suggested change

	!std::is_same<bool, decltype(detect(std::declval<T>()))>::value;
	!std::is_same<decltype(detect(std::declval<T>())),bool>::value;

[bangerth]

Suggested change

	std::declval<U const>().Tstep(std::declval<VectorType &>(),
	std::declval<const U>().Tstep(std::declval<VectorType &>(),

[bangerth]

Suggested change

	!std::is_same<bool, decltype(detect(std::declval<T>()))>::value;
	!std::is_same<decltype(detect(std::declval<T>())),bool>::value;

[bangerth]

What do you do if the function, for whatever reason, returns bool? I suspect that's not very likely, but a common trick is to make the other function return an object of type X where X is a locally declared struct type. No member function of an external class can reasonably return such an X.

[masterleinad]

Looks good. It's time we get to use C++20 concepts :-)

We could just roll our own is_detected for now.

[bangerth]

I was thinking of requires clauses and proper concepts.

How does is_detected work?

[peterrum]

@bangerth I used the following code as template:

dealii/include/deal.II/matrix_free/type_traits.h

Lines 215 to 241 in ad71f79

	// similar to type traits in FEEvaluation, below we add type-traits
	// to distinguish between vectors that provide different interface for
	// operations like update_ghost_values(), compress(), etc.
	// see internal::has_local_element in fe_evaluation.h that documents
	// how those type traits work.
	// a helper type-trait that leverage SFINAE to figure out if type T has
	// void T::update_ghost_values_start(const uint) const
	template <typename T>
	struct has_update_ghost_values_start
	{
	private:
	static bool
	detect(...);
	template <typename U>
	static decltype(std::declval<U const>().update_ghost_values_start(0))
	detect(const U &);
	public:
	static const bool value =
	!std::is_same<bool, decltype(detect(std::declval<T>()))>::value;
	};
	// We need to have a separate declaration for static const members
	template <typename T>
	const bool has_update_ghost_values_start<T>::value;

I don't know all details SFINAE. I don't know why these lines were written like this. Maybe we could proceed with this PR like it is and postpone a possible clean up to later: there are some function in matrix_free/type_traits.h and in lac/precondition.h like this.

[peterrum]

I assume you're going to test this in a follow-up patch somewhere?

Yes.

[bangerth]

Yes, if there is existing art that does it the same way, let's stick with it!

[masterleinad]

I was thinking of requires clauses and proper concepts.

How does is_detected work?

It pretty much does exactly what we are doing here but standardized (and shorter), see https://en.cppreference.com/w/cpp/experimental/is_detected. It might never hit the standard since there are concepts in C++20. Since it will take a good amount of time before requiring C++20 using a hand-rolled is_detected might be a good cleanup in the meantime.

[bangerth]

I would accept a patch based on this.