Runtime 'SIMD' Dispatch Templates for 'C' Code in 'R' Packages

Provides templates and a working example for runtime Single Instruction Multiple Data ('SIMD') dispatch in 'C' code used by 'R' packages. Packages can stage scalar and architecture-specific kernel objects during configuration, then select a compiled and CPU-supported implementation at runtime through guarded function pointers. The package also vendors the header-only 'SIMDe' library for downstream packages through the 'LinkingTo' field.

RsimdDispatch

Pure-C runtime SIMD dispatch templates for R packages. Stage scalar, SSE, AVX, AVX-512, and NEON kernel objects during configuration; link one shared library and switch among compiled and CPU-supported implementations at runtime without unsafe overrides.

Install

install.packages(
  "RsimdDispatch",
  repos = c("https://sounkou-bioinfo.r-universe.dev", "https://cloud.r-project.org")
)

From GitHub:

# install.packages("remotes")
remotes::install_github("sounkou-bioinfo/RsimdDispatch")

Use in other packages

Copy the C dispatch scaffold from your package root:

RsimdDispatch::use_simd_dispatch(pkg = "MyPackage", prefix = "mypkg")

The helper updates DESCRIPTION with LinkingTo: RsimdDispatch, writes the scaffold files, and substitutes package-specific registration and C symbol prefixes. It does not add a runtime dependency on RsimdDispatch. Replace the demo count_nonzero() kernels under tools/kernels/ with your own scalar/SSE/AVX/NEON kernels. The generated configure script probes compiler support, stages selected kernel objects in src/rsd-kernels/, and leaves src/Makevars to link those objects with the baseline R API, CPU detection, and dispatcher.

See the package vignettes for the downstream-package workflow and dispatch semantics.

API

library(RsimdDispatch)

x <- as.raw(c(0, 1, 2, 0, 255))

count_nonzero(x)
#> [1] 3
simd_backend()
#> [1] "avx2"
simd_set_backend("scalar")
simd_backend()
#> [1] "scalar"
simd_set_backend("avx2")
simd_backend()
#> [1] "avx2"
simd_set_backend("auto")
simd_info()[c("compiled_backends", "cpu_supported_backends", "available_backends", "simde_native_backends")]
#> $compiled_backends
#> [1] "scalar" "sse2"   "sse41"  "avx2"   "avx512"
#> 
#> $cpu_supported_backends
#> [1] "scalar" "sse2"   "sse41"  "avx2"  
#> 
#> $available_backends
#> [1] "scalar" "sse2"   "sse41"  "avx2"  
#> 
#> $simde_native_backends
#> [1] "sse2"   "sse41"  "avx2"   "avx512"
simde_info()[c("version", "commit")]
#> $version
#> [1] "0.8.4"
#> 
#> $commit
#> [1] "f3e8262173b7089db9a9d57a9ecef8dd07ad9c97"
simd_dispatch_template_path()
#> [1] "/usr/local/lib/R/site-library/RsimdDispatch/templates/dispatch-c"

simd_set_backend() rejects uncompiled or CPU-unsupported backends. There is no unsafe override.

Example

count_nonzero(x)
#> [1] 3
simd_info()[c("selected_backend", "available_backends")]
#> $selected_backend
#> [1] "avx2"
#> 
#> $available_backends
#> [1] "scalar" "sse2"   "sse41"  "avx2"

Switching backends is same-process because all variants are linked into one R shared library and selected through guarded function pointers.

simd_set_backend("scalar")
scalar_count <- count_nonzero(x)

simd_set_backend("avx2")
avx2_count <- count_nonzero(x)

simd_set_backend("auto")
data.frame(
  scalar = scalar_count,
  avx2 = avx2_count,
  selected_after_auto = simd_backend()
)
#>   scalar avx2 selected_after_auto
#> 1      3    3                avx2

Scalar vs AVX2 benchmark

set.seed(1)
x <- as.raw(sample.int(256L, 50 * 1024 * 1024, replace = TRUE) - 1L)

mark <- bench::mark(
  scalar = {
    simd_set_backend("scalar")
    count_nonzero(x)
  },
  avx2 = {
    simd_set_backend("avx2")
    count_nonzero(x)
  },
  iterations = 20,
  check = TRUE,
  memory = FALSE
)

simd_set_backend("auto")

bench <- data.frame(
  backend = as.character(mark$expression),
  median_ms = as.numeric(mark$median) * 1000,
  mb_per_second = length(x) * as.numeric(mark$`itr/sec`) / 1e6,
  iterations = mark$n_itr,
  row.names = NULL
)
bench$speedup_vs_scalar <- bench$mb_per_second / bench$mb_per_second[bench$backend == "scalar"]

knitr::kable(bench, digits = 3)

Development

Updating bundled SIMDe

This is a maintainer workflow for RsimdDispatch itself, not something downstream packages need to run. Downstream packages use LinkingTo: RsimdDispatch and the copied dispatch template; they do not re-vendor SIMDe.

RsimdDispatch vendors the full SIMDe header tree under inst/include/simde. To update the pinned checkout and regenerate bundled-code notices:

Rscript tools/vendor-simde.R
Rscript tools/update-authors.R

Provenance lives in inst/vendor/simde/VERSION, inst/AUTHORS, and inst/LICENCE.note.

Package maintenance

The top-level Makefile uses GNU make.

make rd       # roxygen docs and NAMESPACE
make readme   # evaluate README.Rmd and write README.md
make test
make check