Home GnuPG

Add 64-bit ARMv8/CE PMULL implementation of CRC

Description

Add 64-bit ARMv8/CE PMULL implementation of CRC

* cipher/Makefile.am: Add 'crc-armv8-ce.c' and
'crc-armv8-aarch64-ce.S'.
* cipher/asm-common-aarch64.h [HAVE_GCC_ASM_CFI_DIRECTIVES]: Add CFI
helper macros.
* cipher/crc-armv8-aarch64-ce.S: New.
* cipher/crc-armv8-ce.c: New.
* cipher/crc.c (USE_ARM_PMULL): New.
(CRC_CONTEXT) [USE_ARM_PMULL]: Add 'use_pmull'.
[USE_ARM_PMULL] (_gcry_crc32_armv8_ce_pmull)
(_gcry_crc24rfc2440_armv8_ce_pmull): New prototypes.
(crc32_init, crc32rfc1510_init, crc24rfc2440_init): Enable ARM PMULL
implementations if supported by HW features.
(crc32_write, crc24rfc2440_write) [USE_ARM_PMULL]: Use ARM PMULL
implementations if enabled.
* configure.ac: Add 'crc-armv8-ce.lo' and 'crc-armv8-aarch64-ce.lo'.

Benchmark on Cortex-A53 (at 1104 Mhz):

Before:

|  nanosecs/byte   mebibytes/sec   cycles/byte

CRC32 | 2.89 ns/B 330.2 MiB/s 3.19 c/B
CRC32RFC1510 | 2.89 ns/B 330.2 MiB/s 3.19 c/B
CRC24RFC2440 | 2.72 ns/B 350.8 MiB/s 3.00 c/B

After (crc32 ~8.4x faster, crc24 ~6.8x faster):

|  nanosecs/byte   mebibytes/sec   cycles/byte

CRC32 | 0.341 ns/B 2796 MiB/s 0.377 c/B
CRC32RFC1510 | 0.342 ns/B 2792 MiB/s 0.377 c/B
CRC24RFC2440 | 0.398 ns/B 2396 MiB/s 0.439 c/B

  • Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>

Details

Provenance
jukiviliAuthored on Apr 26 2019, 6:28 PM
Parents
rCb878a986f3ab: mpi: make stack unwinding work at i386 mpi functions
Branches
Unknown
Tags
Unknown