VPMSUMD acceleration for GCM mode on PPC
440332532a1c
Actions

Description

VPMSUMD acceleration for GCM mode on PPC

* cipher/Makefile.am: Add 'cipher-gcm-ppc.c'.
* cipher/cipher-gcm-ppc.c: New.
* cipher/cipher-gcm.c [GCM_USE_PPC_VPMSUM] (_gcry_ghash_setup_ppc_vpmsum)
(_gcry_ghash_ppc_vpmsum, ghash_setup_ppc_vpsum, ghash_ppc_vpmsum): New.
(setupM) [GCM_USE_PPC_VPMSUM]: Select ppc-vpmsum implementation if
HW feature "ppc-vcrypto" is available.
* cipher/cipher-internal.h (GCM_USE_PPC_VPMSUM): New.
(gcry_cipher_handle): Move 'ghash_fn' at end of 'gcm' block to align
'gcm_table' to 16 bytes.
* configure.ac: Add 'cipher-gcm-ppc.lo'.
* tests/basic.c (_check_gcm_cipher): New AES256 test vector.
* AUTHORS: Add 'CRYPTOGAMS'.
* LICENSES: Add original license to 3-clause-BSD section.

https://dev.gnupg.org/D501:

10-20X speed.

However this Power 9 machine is faster than the last Power 9 benchmarks
on the optimized versions, so while better than the last patch, it is
not all due to the code.

Before:

GCM enc | 4.23 ns/B 225.3 MiB/s - c/B
GCM dec | 3.58 ns/B 266.2 MiB/s - c/B
GCM auth | 3.34 ns/B 285.3 MiB/s - c/B

After:

GCM enc | 0.370 ns/B 2578 MiB/s - c/B
GCM dec | 0.371 ns/B 2571 MiB/s - c/B
GCM auth | 0.159 ns/B 6003 MiB/s - c/B

[jk: coding style fixes, Makefile.am integration, patch from Differential
to git, commit changelog, fixed few compiler warnings]