mmcloughlin/asm

Go library providing algorithms optimized to leverage the characteristics of modern CPUs

asm

Go library providing algorithms optimized to leverage the characteristics of modern CPUs.

Motivation

With the development of Cloud technologies, access to large scale compute capacity has never been easier, and running distributed systems deployed across dozens or sometimes hundreds of CPUs has become common practice. As a side effect of being provided seemingly unlimited (but somewhat expensive) compute capacity, software engineers are now in direct connections with the economical and environmental impact of running the software they develop in production; performance and efficiency of our programs matters today more than it has ever before.

Modern CPUs are complex machines with performance characteristic that may vary by orders of magnitude depending on how they are used. Features like branch prediction, instruction reordering, pipelining, or caching are all input variables that determine the compute throughput that a CPU can achieve. While compilers keep being improved, and often employ micro-optimizations that would be counter-productive for human developers to be responsible for, there are limitations to what they can do, and Assembly still has a role to play in optimizing algorithms on hot code paths of large scale applications.

SIMD instruction sets offer interesting opportunities for software engineers. Taking advantage of these instructions often requires rethinking how the program represents and manipulates data, which is beyond the realm of optimizations that can be implemented by a compiler. When renting CPU time from a Cloud provider, programs that fail to leverage the full sets of instructions available are therefore paying for features they do not use.

This package aims to provide such algorithms, optimized to leverage advanced instruction sets of modern CPUs to maximize throughput and take the best advantage of the available compute power. Users of the package will find functions that have often been designed to work on arrays of values, which is where SIMD and branchless algorithms shine.

The functions in this library have been used in high throughput production environments at Segment, we hope that they will be useful to other developers using Go in performance-sensitive software.

Usage

The library is composed of multiple Go packages intended to act as logical groups of functions sharing similar properties:

Package	Purpose
ascii	library of functions designed to work on ASCII inputs
base64	standard library compatible base64 encodings
bswap	byte swapping algorithms working on arrays of fixed-size items
cpu	definition of the ABI used to detect CPU features
mem	functions operating on byte arrays
qsort	quick-sort implementations for arrays of fixed-size items
slices	functions performing computations on pairs of slices
sortedset	functions working on sorted arrays of fixed-size items

When no assembly version of a function is available for the target platform, the package provides a generic implementation in Go which is automatically picked up by the compiler.

Showcase

The purpose of this library being to improve the runtime efficiency of Go programs, we compiled a few snapshots of benchmark runs to showcase the kind of improvements that these code paths can expect from leveraging SIMD and branchless optimizations:

goos: darwin
goarch: amd64
cpu: Intel(R) Core(TM) i9-8950HK CPU @ 2.90GHz

pkg: github.com/segmentio/asm/ascii
name                  old time/op    new time/op     delta
EqualFoldString/0512     276ns ± 1%       21ns ± 2%    -92.50%  (p=0.008 n=5+5)

name                  old speed      new speed       delta
EqualFoldString/0512  3.71GB/s ± 1%  49.44GB/s ± 2%  +1232.79%  (p=0.008 n=5+5)

pkg: github.com/segmentio/asm/bswap
name    old time/op    new time/op     delta
Swap64    11.2µs ± 1%      0.9µs ± 9%    -92.06%  (p=0.008 n=5+5)

name    old speed      new speed       delta
Swap64  5.83GB/s ± 1%  73.67GB/s ± 9%  +1162.98%  (p=0.008 n=5+5)

pkg: github.com/segmentio/asm/qsort
name            old time/op    new time/op     delta
Sort16/1000000     269ms ± 2%       46ms ± 3%   -83.08%  (p=0.008 n=5+5)

name            old speed      new speed       delta
Sort16/1000000  59.4MB/s ± 2%  351.2MB/s ± 3%  +491.24%  (p=0.008 n=5+5)

Maintenance

Generation of the assembly code is managed with AVO, and orchestrated by a Makefile which helps maintainers rebuild the assembly source code when the AVO files are modified.

The repository contains two Go modules; the main module is declared as github.com/segmentio/asm at the root of the repository, and the second module is found in the build subdirectory.

The build module is used to isolate build dependencies from programs that import the main module. Through this mechanism, AVO does not become a dependency of programs using github.com/segmentio/asm, keeping the dependency management overhead minimal for the users, and allowing maintainers to make modifications to the build package.

Versioning of the two modules is managed independently; while we aim to provide stable APIs on the main package, breaking changes may be introduced on the build package more often, as it is intended to be ground for more experimental constructs in the project.

purego

Programs in the build module should add the following declaration:

func init() {
	ConstraintExpr("!purego")
}

It instructs AVO to inject the !purego tag in the generated files, allowing compilation of the libraries without any assembly optimizations with a build command such as:

go build -tags purego ...

This is mainly useful to compare the impact of using the assembly optimized versions instead of the simpler Go-only implementations.