Any plans for a ttmath backend?

[LegalizeAdulthood]

I only stumbled across ttmath today. It was mentioned in the docs for Boost.Geometry.

Maybe if there aren't any plans to implement this directly as a backend for Boost.Multiprecision it could be used as a documentation example of how to add your own backend?

[ckormanyos]

Hi Richard (@LegalizeAdulthood) this is a good idea. I've wrapped a couple of my own backends for Multiprecision.

Generally, I distinguish between wrapping one for local or project-specific work as opposed to actually getting it formally approved and tested for Boost.

The former is rather straghtforward. Whereby the latter can be tougher. We just finished getting cpp_double_fp_backend into MP and it was quite an effort. But just wrapping a backend is much, much more streamlined.

You are right, we do need a real world nuts-and-bolts tutorial for this. Let's talk about this as we evolve this project.

Cc: @jzmaddock and @mborland

[LegalizeAdulthood]

You are right, we do need a real world nuts-and-bolts tutorial for this. Let's talk about this as we evolve this project.

That would be great; I'm giving a presentation on Boost.Multiprecision in a couple weeks, I'll give it a go in my example.

[ckormanyos]

Very cool. Thanks for getting into Boost and Multiprecision.

So I was just looking at the new work on cpp_double_fp_backend. If you look at about the first $150$ lines of cpp_double_fp.hpp, you will see a series of functions called eval_whatever.

Most of these are mandatory for a floating-point backend, with exceptions like eval_exp, eval_log and eval_pow (I did these for enhanced performance).

Basically if you make your eval_whatever functions and get the right headers, you plug into Multiprecison and Math.

You must also make two or three tuples and clarify your number category and also make a kind of precision structure. The precision is near the bottom of the file. And the tuples for built-in types are near the top of the class (signed_types and unsigned_types and float_types). The types are needed for the internals of the backend.

[ckormanyos]

There is another completely unrelated yet interesting example in ckormanyos/mp_cpp. This project is real legacy and looks a little bit like cpp_dec_float. I worked on this as long ago as in the 90s. It's an old toy. It would not pass all of the Multiprecision testing without significant work. There is even a bug somewhere in millions of digits.

Again, it is more of a toy. Over there (and also in the wide-decimal project), I am trying my hand at Karatsuba and Toom-Cook and FFT multiplication.

I wrapped the legacy mp_cpp class as a Multiprecision backend here.

[jzmaddock]

There are (admittedly rather minimal) docs on adding a new backend here: https://www.boost.org/doc/libs/latest/libs/multiprecision/doc/html/boost_multiprecision/tut/new_backend.html

More importantly is https://www.boost.org/doc/libs/latest/libs/multiprecision/test/skeleton_backend.hpp which basically does a lot of the drudgery for you and lets you "fill in the blanks". The list of required functions is surprisingly short.

I might be inclined to use the Intel Decimal Library as a case study as a new backend if we were going to put the work in for one.

[ckormanyos]

More importantly is ... [skeleton].

Yes thanks John. I have used skeleton on several occasions to get started. It is the right starting point.

inclined to use the Intel Decimal Library

We have a couple of pending client requests for backend support. If you go through the open issus in Multiprecision, you can find them.

Some of our clients want quad-double, and I think I will start on that in a few months. Then there is a package with the terse name arb. I think there was interest in that one too. I read an article once about a world-record-breaking calculation of the non-trivial zeros of the complex-valued Riemann-zeta function. And they used a package for interval-math there. I forgot its name.

[ckormanyos]

And don't forget to specify std::numeric_limits for your type. This is an interesting case, since we extend the std-namespace. I think it is allowed, but it might be a topic of debate in some circles.

[ckormanyos]

The issue for arb is here.

In addition, I think that arb was actually used in "The Riemann hypothesis is true up to $10^{12}$ ...". I forgot the correct citation but it was intruiging.

It's interval math, so maybe not what you're looking for. But there was interest in that wrapper.