[openpgp] SHA-x performance (was: SHA3 algorithm ids)

[Werner Koch <[e-mail]>]

On Tue, 11 Aug 2015 18:39, [e-mail] said:

> A huge number of devices, and in particular ones with less CPU power, are
> still 32-bit, and will remain so for a long time, probably more or less
> indefinitely.

Does anyone know a summary of SHA-256 performance on standard CPUs
with dedicated SHA hardware?  The Padlock engine has this but I don't
know whether other CPUs also provide hardware support.  I assume that on
x86 the AVX instructions are as good as dedicated support:

FWIW, Libgcrypt on an i5-2410M (64 bit) gives this:

                |  nanosecs/byte   mebibytes/sec   cycles/byte
 SHA1           |      1.92 ns/B     496.5 MiB/s      4.42 c/B
 SHA256         |      4.42 ns/B     215.6 MiB/s     10.17 c/B
 SHA512         |      2.97 ns/B     321.1 MiB/s      6.83 c/B

with AVX disabled:

 SHA1           |      2.27 ns/B     419.7 MiB/s      5.23 c/B
 SHA256         |      5.26 ns/B     181.1 MiB/s     12.11 c/B
 SHA512         |      3.61 ns/B     264.0 MiB/s      8.31 c/B

with AVX and SSSE3 disabled:

 SHA1           |      3.27 ns/B     292.0 MiB/s      7.51 c/B
 SHA256         |      7.50 ns/B     127.1 MiB/s     17.26 c/B
 SHA512         |      4.68 ns/B     203.6 MiB/s     10.78 c/B

We have no optimized SHA3 yet; for reference here are the numbers from
the unoptimized version:

 SHA3-256       |      5.70 ns/B     167.3 MiB/s     13.11 c/B
 SHA3-512       |     10.66 ns/B     89.46 MiB/s     24.52 c/B


Salam-Shalom,

   Werner

-- 
Die Gedanken sind frei.  Ausnahmen regelt ein Bundesgesetz.