1528113 - Use ARM's crypt extension for SHA256

Status: RESOLVED FIXED

(NSS :: Libraries, enhancement, P3)

Description

[Makoto Kato [:m_kato]]

When I implement it as test, it improves a lot of performance for SHA256.

• ThunderX is 3x faster then C implementation on ARM64
• Cortex-A53 is 8x faster then C implementation on ARM64

For Android/aarch64 (and Windows), we should use it.

Comment 1

[Makoto Kato [:m_kato]]

result on a1.xlarge on AWS (Cortex-A72)

      mode          in  opreps  cxreps     context          op   time(sec)     thrgput
  sha256_e         8Gb      5M       0       0.000   10000.000      10.000       842Mb <-- with patch
  sha256_e         1Gb    943T       0       0.000   10000.000      10.000       154Mb <-- without patch

Comment 2

[Makoto Kato [:m_kato]]

https://treeherder.mozilla.org/#/jobs?repo=nss-try&revision=c2004a8b9b590d44fb8644dfb1e073c629fa20af

Comment 3

[Makoto Kato [:m_kato]]

Attachment: Bug 1528113 - Use ARM's crypto extension for SHA256

ARMv8 CPU has accelerated hardware instruction for SHA256 that supports GCC 4.9+. We should use it if available.

Comment 4

[Makoto Kato [:m_kato]]

Attachment: Bug 1528113 - Use ARM's crypt extension for SHA1

Comment 5

[Val Packett]

AES has landed (again) so this needs rebase.

(btw, tested on A72, works great)

Comment 6

[Makoto Kato [:m_kato]]

I will file a new bug for SHA1's implementation.

Comment 7

[Kevin Jacobs [:kjacobs]]

https://hg.mozilla.org/projects/nss/rev/61c83f79e90c4fdd159f0ee12d559f26c5f46f74