awscrt

R bindings to the CRC routines of the AWS Common Runtime: CRC32, CRC32C and CRC64NVME, hardware-accelerated where the CPU allows it, with a portable software fallback. Output is byte-for-byte what Python’s awscrt.checksums produces (as big-endian bytes), which is what S3 flexible checksums use.

Only aws-checksums (and the small part of aws-c-common it needs) is vendored. There is no HTTP, TLS, event loop or threading, and nothing is downloaded at install time.

Installation

# install.packages("pak")
pak::pak("paws-r/aws-crt-r")

Usage:

library(awscrt)

crc32c("123456789")               # "e3069283"
crc64nvme("123456789")            # "ae8b14860a799888"
crc32c("123456789", raw = TRUE)   # raw vector, big-endian: e3 06 92 83

# Incremental hashing: pass the previous (raw) result back in
a <- crc32c(charToRaw("1234"), raw = TRUE)
identical(crc32c(charToRaw("56789"), a), crc32c("123456789"))

crc32c_file("big.bin")            # chunked read, constant memory
cpu_features()                    # what the runtime dispatch can use

Results are hexadecimal strings by default, because a 64-bit CRC does not fit in an R double. Use raw = TRUE for the big-endian raw vector, which is also the form previous expects. Base64-encoding for HTTP headers is left to the caller.

Build notes

./configure probes the compiler for SSE4.2 / PCLMULQDQ / AVX-512 (x86-64) or CRC / PMULL (arm64), writes src/generated/aws/common/config.h and src/Makevars, and applies the SIMD flags only to the files that need them. CPU support is still checked at run time. Windows (MinGW) always uses the software path, as upstream does. To force a software-only build:

R CMD INSTALL --configure-vars='AWSCRT_SOFTWARE_ONLY=1' .

At run time options(awscrt.force_software = TRUE) bypasses hardware dispatch.

Refresh the vendored sources with tools/update-vendor.sh.