[Proposal] Start a project around various numeric types

[tannergooding]

We have a number of proposals on CoreFX around exposing various new numeric types. It would likely be a good idea to have a project here to prototype some of these.

Some of the ideas that I think would likely be applicable are the following:

Types which are generally considered ABI primitives:

• The IEEE binary128 (quad-precision) floating-point type
• The IEEE decimal32, decimal64, and decimal128 floating-point types
• The Int128 and UInt128 types

Types which are fairly core to machine-learning and graphics scenarios:

• The IEEE binary16 (half-precision) floating-point type
• The bfloat16 floating-point type

Types which allow arbitrary precision arithmetic:

• The IEEE binary{k} (where k >= 128 and (k % 32) == 0) floating-point type (effectively a BigFloat type)
• The IEEE decimal{k} (where k >= 32 and (k % 32) == 0) floating-point type (effectively a BigDecimal type)

[tannergooding]

CC. @danmosemsft

[tannergooding]

Some of the existing issues requesting these or similar types are:

• https://github.com/dotnet/corefx/issues/17147
• https://github.com/dotnet/corefx/issues/17267
• https://github.com/dotnet/corefx/issues/25702
• https://github.com/dotnet/corefx/issues/27033

[AlgorithmsAreCool]

This is cool, I just had to roll my own BigRational to support exact precision decimal arithmetic for a project.

Would you consider mutable arbitrary precision numbers?
I recall the original discussions around wanting to keep the semantics of BigInteger the same as normal ints.

But when you are doing alot of calculations the copying becomes a big (heh) pain.

It would be really nice to have a BigIntegerBuilder class or something that could accumulate repeated updates and only reallocate buffers when it was actually needed (or rent buffers in smaller segments).

I can provide a more concrete use case if you are open to the concept.

[fanoI]

Happy to see this moving forward.

Can I hope the implementations of these new numeric types will be all in managed code?

[tannergooding]

Can I hope the implementations of these new numeric types will be all in managed code?

Yes. Although they likely will take advantage of the new hardware intrinsics feature, where applicable.

[tannergooding]

@AlgorithmsAreCool, I think having the various arbitrary precision types and corresponding Builder types is probably the correct way forward.

This is a pattern that has been used with fairly good success for other data types and BigInteger has already shipped as readonly, so it would likely be good to keep things consistent moving forward.

[Happypig375]

If anyone needs this right now, there is https://github.com/peteroupc/Numbers which implements all of this via configurable context settings.