Optimize AES-XTS mode in OpenSSL for aarch64
authorXiaokangQian <xiaokang.qian@arm.com>
Fri, 13 Mar 2020 03:27:34 +0000 (03:27 +0000)
committerTomas Mraz <tmraz@fedoraproject.org>
Thu, 12 Nov 2020 10:09:22 +0000 (11:09 +0100)
commit9ce8e0d17e608de4f85f7543c52b146e3c6a2291
tree6b7472519cee7aa202123ed90c9b982be41457a5
parentc87a7f31a3db97376d764583ad5ee4a76db2cbef
Optimize AES-XTS mode in OpenSSL for aarch64

Aes-xts mode can be optimized by interleaving cipher operation on
several blocks and loop unrolling. Interleaving needs one ideal
unrolling factor, here we adopt the same factor with aes-cbc,
which is described as below:
If blocks number > 5, select 5 blocks as one iteration,every
loop, decrease the blocks number by 5.
If left blocks < 5, treat them as tail blocks.
Detailed implementation has a little adjustment for squeezing
code space.
With this way, for small size such as 16 bytes, the performance is
similar as before, but for big size such as 16k bytes, the performance
improves a lot, even reaches to 2x uplift, for some arches such as A57,
the improvement even reaches more than 2x uplift. We collect many
performance datas on different micro-archs such as thunderx2,
ampere-emag, a72, a75, a57, a53 and N1, all of which reach 0.5-2x uplift.
The following table lists the encryption performance data on aarch64,
take a72, a75, a57, a53 and N1 as examples. Performance value takes the
unit of cycles per byte, takes the format as comparision of values.
List them as below:

A72:
                            Before optimization     After optimization  Improve
evp-aes-128-xts@16          8.899913518             5.949087263         49.60%
evp-aes-128-xts@64          4.525512668             3.389141845         33.53%
evp-aes-128-xts@256         3.502906908             1.633573479         114.43%
evp-aes-128-xts@1024        3.174210419             1.155952639         174.60%
evp-aes-128-xts@8192        3.053019303             1.028134888         196.95%
evp-aes-128-xts@16384       3.025292462             1.02021169          196.54%
evp-aes-256-xts@16          9.971105023             6.754233758         47.63%
evp-aes-256-xts@64          4.931479093             3.786527393         30.24%
evp-aes-256-xts@256         3.746788153             1.943975947         92.74%
evp-aes-256-xts@1024        3.401743802             1.477394648         130.25%
evp-aes-256-xts@8192        3.278769327             1.32950421          146.62%
evp-aes-256-xts@16384       3.27093296              1.325276257         146.81%

A75:
                            Before optimization     After optimization  Improve
evp-aes-128-xts@16          8.397965173             5.126839098         63.80%
evp-aes-128-xts@64          4.176860631             2.59817764          60.76%
evp-aes-128-xts@256         3.069126585             1.284561028         138.92%
evp-aes-128-xts@1024        2.805962699             0.932754655         200.83%
evp-aes-128-xts@8192        2.725820131             0.829820397         228.48%
evp-aes-128-xts@16384       2.71521905              0.823251591         229.82%
evp-aes-256-xts@16          11.24790935             7.383914448         52.33%
evp-aes-256-xts@64          5.294128847             3.048641998         73.66%
evp-aes-256-xts@256         3.861649617             1.570359905         145.91%
evp-aes-256-xts@1024        3.537646797             1.200493533         194.68%
evp-aes-256-xts@8192        3.435353012             1.085345319         216.52%
evp-aes-256-xts@16384       3.437952563             1.097963822         213.12%

A57:
                            Before optimization     After optimization  Improve
evp-aes-128-xts@16          10.57455446             7.165438012         47.58%
evp-aes-128-xts@64          5.418185447             3.721241202         45.60%
evp-aes-128-xts@256         3.855184592             1.747145379         120.66%
evp-aes-128-xts@1024        3.477199757             1.253049735         177.50%
evp-aes-128-xts@8192        3.36768104              1.091943159         208.41%
evp-aes-128-xts@16384       3.360373443             1.088942789         208.59%
evp-aes-256-xts@16          12.54559459             8.745489036         43.45%
evp-aes-256-xts@64          6.542808937             4.326387568         51.23%
evp-aes-256-xts@256         4.62668822              2.119908754         118.25%
evp-aes-256-xts@1024        4.161716505             1.557335554         167.23%
evp-aes-256-xts@8192        4.032462227             1.377749511         192.68%
evp-aes-256-xts@16384       4.023293877             1.371558933         193.34%

A53:
                            Before optimization     After optimization  Improve
evp-aes-128-xts@16          18.07842135             13.96980808         29.40%
evp-aes-128-xts@64          7.933818397             6.07159276          30.70%
evp-aes-128-xts@256         5.264604704             2.611155744         101.60%
evp-aes-128-xts@1024        4.606660117             1.722713454         167.40%
evp-aes-128-xts@8192        4.405160115             1.454379201         202.90%
evp-aes-128-xts@16384       4.401592028             1.442279392         205.20%
evp-aes-256-xts@16          20.07084054             16.00803726         25.40%
evp-aes-256-xts@64          9.192647294             6.883876732         33.50%
evp-aes-256-xts@256         6.336143161             3.108140452         103.90%
evp-aes-256-xts@1024        5.62502952              2.097960651         168.10%
evp-aes-256-xts@8192        5.412085608             1.807294191         199.50%
evp-aes-256-xts@16384       5.403062591             1.790135764         201.80%

N1:
                            Before optimization     After optimization  Improve
evp-aes-128-xts@16          6.48147613              4.209415473         53.98%
evp-aes-128-xts@64          2.847744115             1.950757468         45.98%
evp-aes-128-xts@256         2.085711968             1.061903238         96.41%
evp-aes-128-xts@1024        1.842014669             0.798486302         130.69%
evp-aes-128-xts@8192        1.760449052             0.713853939         146.61%
evp-aes-128-xts@16384       1.760763546             0.707702009         148.80%
evp-aes-256-xts@16          7.264142817             5.265970454         37.94%
evp-aes-256-xts@64          3.251356212             2.41176323          34.81%
evp-aes-256-xts@256         2.380488469             1.342095742         77.37%
evp-aes-256-xts@1024        2.08853022              1.041718215         100.49%
evp-aes-256-xts@8192        2.027432668             0.944571334         114.64%
evp-aes-256-xts@16384       2.00740782              0.941991415         113.10%

Add more XTS test cases to cover the cipher stealing mode and cases of different
number of blocks.

CustomizedGitHooks: yes
Change-Id: I93ee31b2575e1413764e27b599af62994deb4c96

Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Tomas Mraz <tmraz@fedoraproject.org>
(Merged from https://github.com/openssl/openssl/pull/11399)
crypto/aes/asm/aesv8-armx.pl
include/crypto/aes_platform.h
test/recipes/30-test_evp_data/evpciph_aes_common.txt