Refine SIMD implementation

[KHLee529]

Abstract

According to the discussion in PR #641, there is an unintended recursive program flow in current SIMD implementation. Recursion is not proper in low level basic container especially for performance-critical functions.
This issue tracks the refinement of SIMD implementation and reconsider the overall structure of our SIMD library.

Tasks

• Remove unintended recursion
• Discuss on SIMD implementation

Problem on current SIMD implementation

Current SIMD implementation is based on functionality module. For example, add, subtract, multiply, divide are implemented in every different SIMD platform. In this form, the logic of each functionality are actually implemented multiple times in different part of the library. Even platform without SIMD support needs its own "generic" version of implementation. Comparing with other SIMD library, take xsimd for instance, are modularized by the SIMD basic instructions such as load value to vector registry, store vector register to memory, arithmetic operations on two vector registers. In this way, the program logic of different SIMD platform can be unify and implement only once, which should be more maintainable and understandable.

[KHLee529]

@yungyuc @nccvector @tigercosmos @tychuang1211 this issue is opened for the follow-up discussion of PR #641. Please share your opinion.

[tigercosmos]

I agree with (1) and (2). However, for the SIMD API desgin, it's somehow hard to design now because we don't have actual use case yet.

[KHLee529]

I agree with (1) and (2). However, for the SIMD API desgin, it's somehow hard to design now because we don't have actual use case yet.

I don't really get it, what are the (1) and (2) you mentioned?

[tigercosmos]

Ops, sorry, it means the task 1 and task 2 in the list.

[yungyuc]

• Remove unintended recursion
• Discuss on SIMD implementation

Did you mean task 1 and task 2 above?

Ops, sorry, it means the task 1 and task 2 in the list.

It will be easier to describe by quoting the parts.