SLP - Superword Level Parallelism

gcc에서 vectorize  옵션을 통해 빌드 하는데

-fopt-info-vec-all -O3 -mavx

 

메시지들 중에서 SLP 라는 용어 발견

note: ===vect_slp_analyze_bb=== note: Final SLP tree for instance: note: Build SLP for varname = 900; note: Build SLP failed: unrolling required in basic block SLP note: vectorizing stmts using SLP.

 

gcc 메뉴얼에도 SLP 어쩌구는 나오는데 약어가 안나온다.

2011-10-23 Vectorization of reduction in loop SLP. Both multiple reduction cycles and reduction chains are supported. Various basic block vectorization (SLP) improvements, such as better data dependence analysis, support of misaligned accesses and multiple types, cost model.

[링크 : https://gcc.gnu.org/projects/tree-ssa/vectorization.html#using]

 

Superword-Level Parallelism (SLP) vectorizer

[링크 : https://rcor.me/papers/cgo19snslp.pdf]

 

Superword level parallelism (SLP) is an advanced method of traditional vectorization that facilitates parallelism across loop iterations, SIMD, and basic blocks. Vectorization in this context is the process used to complete several similar tasks (or instructions) simultaneously, therefore saving processing time and resources.

[링크 : https://www.webopedia.com/definitions/superword-level-parallelism/]