Improved AES performance

[Werner Koch]

Hi!

I hacked libgcrypt a bit to see how to increase the symmetrical cipher
performance.  This should benefit all architecture except for Padlock,
which I have not yet changed.  

The current SVN has the changes for AES in CFB and CBC mode.  On my box
it gives about 25% speedup with the benchmark tool.  That tool has two
new options:

  --large-buffers 
                  passes larger blocks of data to the encryption function
                  and reduces the loop count instead.

  --cipher-repetitions N  
                  may be used to increase the loop count by a factor N.
                  This is gives more accurate timings on fas machines.

Example usage:

  tests/benchmark --cipher-repetition 10 cipher aes aes192 aes256


I did a real work test with gpg2 on a 2.9GB MPEG file.  First a plain
store operation to see the I/O overhead.

  $ time gpg2 --store -z0 -v --batch --passphrase 'test'\
            <video/Rockpalast-The_Who.mpg >/dev/null
  gpg: writing to stdout
  
  real    1m47.123s
  user    0m5.112s
  sys     0m5.232s

So this gives us a mere 5 seconds.  Now with AES-128 using a stock
libgcrypt 1.4.0:
  
  $ time gpg2 -c --cipher aes  -z0 -v --batch --passphrase 'test' \
            <video/Rockpalast-The_Who.mpg >/dev/null
  gpg: using cipher AES
  gpg: writing to stdout
  
  real    3m58.025s
  user    2m53.814s
  sys     0m5.968s
  
Yields 73 seconds.  Again with the latest libgcrypt:

  $ time gpg2 -c --cipher aes  -z0 -v --batch --passphrase 'test' \
            <video/Rockpalast-The_Who.mpg >/dev/null
  gpg: using cipher AES
  gpg: writing to stdout
  
  real    3m22.304s
  user    2m19.503s
  sys     0m5.852s

Yields 139 seconds.  Comparing the corrected user times

 173 - 5 = 168
 139 - 5 = 134 

shows that encryption is about 25% faster in CFB mode.
  


Salam-Shalom,

   Werner

  
-- 
Die Gedanken sind frei.  Auschnahme regelt ein Bundeschgesetz.