This crate provides a set of bulk homogeneous-slice-typed operations. It's to be used as a building block for collection-oriented evaluators that don't JIT, but still want good performance from a fixed repertoire of type-specialized inner loops. It provides three notable features:
Isolates in one compilation unit the code amplification factor of the cartesian product of all primitive operations across all primitive types, compiled into (where possible) efficient SIMD loops. They only get compiled once, here.
Abstracts away the question of runtime selection of the most efficient implementation (eg. rayon chunking & SIMD, possibly GPU in the future).
Provides "outer" column-vs-column or vs-constant dynamic double-dispatches that are expected to be amortized over the inner loops, as well as the dynamic dispatch from ops to functions. Nothing in this crate should wind up inlined much into its users, despite going very fast once inside!
This is the kind of building block used in eg. APL evaluators or column-oriented databases, where the outer layer is an interpreter and it maintains good-but-not-maximal performance by having all the inner layers go fast: the homogeneous-slice-typed operators here. For data intensive programs it's a simple strategy that seems to pay off well, since the outer interpreter loop is only returned to relatively infrequently, and the code footprint of these inner loops is compiled once into the interpreter, so constant across all user programs and relatively cache-friendly.
T. Kersten, V. Leis, A. Kemper, T. Neumann, A. Pavlo, P. Boncz. Everything You Always Wanted to Know About Compiled and Vectorized Queries But Were Afraid to Ask. PVLDB, 11 (13): 2209 - 2222, 2018.
DOI: https://doi.org/10.14778/3275366.3275370
http://www.vldb.org/pvldb/vol11/p2209-kersten.pdf
Benchmarks operating on pairs of 16MB columns.
SIMD speedup over scalar loops:
~~~ $ RAYONNUMTHREADS=1 cargo bench test addu32newel ... bench: 1,056,316 ns/iter (+/- 455,605) = 3970 MB/s test addu32scalar ... bench: 2,099,241 ns/iter (+/- 37,008) = 1998 MB/s test addu64newel ... bench: 2,265,895 ns/iter (+/- 300,628) = 3702 MB/s test addu64scalar ... bench: 2,202,050 ns/iter (+/- 84,681) = 3809 MB/s test addu8newel ... bench: 346,318 ns/iter (+/- 58,334) = 3027 MB/s test addu8scalar ... bench: 2,089,541 ns/iter (+/- 52,697) = 501 MB/s test ltu16newel ... bench: 392,897 ns/iter (+/- 74,161) = 5337 MB/s test ltu16scalar ... bench: 1,800,218 ns/iter (+/- 85,009) = 1164 MB/s test ltu32newel ... bench: 636,441 ns/iter (+/- 122,298) = 6590 MB/s test ltu32scalar ... bench: 2,101,893 ns/iter (+/- 63,897) = 1995 MB/s test ltu8newel ... bench: 335,296 ns/iter (+/- 45,917) = 3127 MB/s test ltu8scalar ... bench: 2,105,158 ns/iter (+/- 57,694) = 498 MB/s ~~~
Plus thread-level parallel speedup:
~~~ $ RAYONNUMTHREADS=40 cargo bench test addu32newel ... bench: 163,534 ns/iter (+/- 31,830) = 25647 MB/s test addu32scalar ... bench: 2,112,953 ns/iter (+/- 87,481) = 1985 MB/s test addu64newel ... bench: 283,384 ns/iter (+/- 58,405) = 29601 MB/s test addu64scalar ... bench: 2,254,990 ns/iter (+/- 104,615) = 3720 MB/s test addu8newel ... bench: 78,888 ns/iter (+/- 19,092) = 13291 MB/s test addu8scalar ... bench: 2,099,180 ns/iter (+/- 70,613) = 499 MB/s test ltu16newel ... bench: 99,430 ns/iter (+/- 30,703) = 21091 MB/s test ltu16scalar ... bench: 1,821,691 ns/iter (+/- 58,804) = 1151 MB/s test ltu32newel ... bench: 129,451 ns/iter (+/- 39,019) = 32400 MB/s test ltu32scalar ... bench: 2,102,977 ns/iter (+/- 27,062) = 1994 MB/s test ltu8newel ... bench: 73,589 ns/iter (+/- 20,625) = 14249 MB/s test ltu8scalar ... bench: 2,114,156 ns/iter (+/- 82,354) = 495 MB/s ~~~
Wikipedia:
A newel, also called a central pole or support column, is the central supporting pillar of a staircase.