1592869 - Use NEON for ctr_xor

Status: RESOLVED FIXED

(NSS :: Libraries, enhancement, P1)

Description

[Makoto Kato [:m_kato]]

This bench is on Amazon's arm64 instance (Cortex-A72). Using NEON for ctr_xor, aes_ctr improves 30%-40%.

Before

 #     mode          in symmkey  opreps  cxreps     context          op   time(sec)     thrgput
 aes_ctr_d         2Gb     256     35M       0       0.000   10000.000      10.000       217Mb

 #     mode          in symmkey  opreps  cxreps     context          op   time(sec)     thrgput
 aes_ctr_e         2Gb     256     35M       0       0.000   10000.000      10.000       218Mb

After

 #     mode          in symmkey  opreps  cxreps     context          op   time(sec)     thrgput
 aes_ctr_d         2Gb     256     48M       0       0.000   10000.000      10.000       294Mb

 #     mode          in symmkey  opreps  cxreps     context          op   time(sec)     thrgput
 aes_ctr_e         2Gb     256     48M       0       0.000   10000.000      10.000       295Mb

Comment 1

[Makoto Kato [:m_kato]]

Android/arm32 build uses NEON as default compiler option, and we don't support non-NEON device on Android, so it can use NEON as default even if Android/arm32.

Comment 2

[Makoto Kato [:m_kato]]

Attachment: Bug 1592869 - Use NEON for ctr_xor.

Using NEON for ctr_xor, aes_ctr can improve 30%-40%i decode/encode time on Cortex-A72.

Comment 3

[J.C. Jones [:jcj] (he/they)]

https://hg.mozilla.org/projects/nss/rev/d244c72879080747cfa4275f04a02969ce0be932