ARMv8 AES encryption, 4x blocks
6 views
Skip to first unread message
Jeffrey Walton
Aug 15, 2017, 12:46:52 PM8/15/17
to Crypto++ Users
Hi Everyone,
I cut-over to a wider AES implementation. Its the implementation that Wei uses for AESNI. Its on my testing fork at the moment. Also see https://github.com/noloader/cryptopp/commit/f2a303c30b5256d3.
The numbers are very good for ECB, CTR and CBC. ECB dropped from 2.3 cpb to 1.0 cpb. CTR dropped from 2.5 cpb to 1.1. And CBC dropped from 3.7 to 2.6 cpb. OFB and CFB stayed about the same.
Testing also shows we don't want to unroll loops when running 4 lanes at a time. Things slowed down by 0.1 cpb.
Jeff
I cut-over to a wider AES implementation. Its the implementation that Wei uses for AESNI. Its on my testing fork at the moment. Also see https://github.com/noloader/cryptopp/commit/f2a303c30b5256d3.
The numbers are very good for ECB, CTR and CBC. ECB dropped from 2.3 cpb to 1.0 cpb. CTR dropped from 2.5 cpb to 1.1. And CBC dropped from 3.7 to 2.6 cpb. OFB and CFB stayed about the same.
Testing also shows we don't want to unroll loops when running 4 lanes at a time. Things slowed down by 0.1 cpb.
Jeff
Reply all
Reply to author
Forward
0 new messages