This directory contains various ciphers and digests pulled out of SSLeay.

There is x86 assember for
rc4, des, blowfish, cast5, md5 and sha1.

On a pentium the md5 takes 337 cycles per block, and is faster than the
speed listed in 'Even Faster Hashing on the Pentium' (345 cycles).
The sha1 inner loop, is the same speed, (837 cycles).

Blowfish has an inner loop that is 9 cycles per round.  There is a
faster version for the pentium pro, but the default version is probably
the best option.

There are 2 variants on the CAST5 implementation.  One has 13 cycles per
round, and the other has 14.  The 13 cycle version unfortunatly runs
%30 slower on a pentium pro than the 14 cycle version.

RC4 processes 8 bytes per 70 cycles.

cbc mode of these ciphers is implemented via assembler, but not inline
code.  If you want another %2-3 speedup, you could easily remove the
function call overhead at the expense of increased code size for the library.

Anyway, what does this mean in the real world?  Using the SSLeay 'speed'
test program, under linux on a pentium 100,

built on Tue Nov  4 02:52:29 EST 1997
options:bn(64,32) md2(int) rc4(ptr,int) des(ptr,risc1,16,long) idea(int) blowfish(ptr2)
C flags:gcc -DL_ENDIAN -DTERMIO -DBN_ASM -O3 -fomit-frame-pointer -m486 -Wall -Wuninitialized -DMD5_ASM -DSHA1_ASM
The 'numbers' are in 1000s of bytes per second processed.
type              8 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes
md5               993.15k     5748.27k    11944.70k    16477.53k    18287.27k
sha1              563.24k     2851.67k     5363.71k     6879.23k     7441.07k
rc4              7876.70k    10400.85k    10825.90k    10943.49k    10745.17k
des cbc          2047.39k     2188.25k     2188.29k     2239.49k     2233.69k
des ede3          660.55k      764.01k      773.55k      779.21k      780.97k
idea cbc          653.93k      708.48k      715.43k      719.87k      720.90k
rc2 cbc           648.08k      702.23k      708.78k      711.00k      709.97k
blowfish cbc     3764.39k     4288.66k     4375.04k     4497.07k     4423.68k
cast cbc         2757.14k     2993.75k     3035.31k     3078.90k     3055.62k

blowfish cbc [*] 3258.81k     3673.47k     3767.30k     3774.12k     3719.17k
cast cbc    [**] 2677.05k     3164.78k     3273.05k     3287.38k     3244.03k

[*]  pentium pro specific version
[**] pentium specific version

For a pentum pro 200, Windows 95, SSLeay with DLLs
built on Tue Nov  4 08:57:30 EST 1997
options:bn(64,32) md2(int) rc4(idx,int) des(idx,cisc,4,long) idea(int) blowfish(ptr2)
C flags:cl /W3 /WX /G5 /Ox /O2 /Ob2 /Gs0 /GF /Gy /nologo -DWIN32 -DL_ENDIAN -DBN_ASM -DMD5_ASM -DSHA1_ASM /MD
The 'numbers' are in 1000s of bytes per second processed.
type              8 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes
md5              2251.85k    11966.63k    22944.77k    29916.58k    32729.64k
sha1             1398.85k     6621.89k    11831.61k    14722.02k    15863.48k
rc4             15744.38k    21239.13k    22093.89k    22419.84k    22564.58k
des cbc          4147.03k     4571.95k     4654.13k     4673.84k     4654.13k
des ede3         1564.92k     1631.00k     1642.47k     1646.21k     1641.86k
idea cbc         2582.06k     2888.64k     2936.78k     2953.74k     2949.58k
rc2 cbc          1646.37k     1782.23k     1800.59k     1805.24k     1806.96k
blowfish cbc     6052.39k     7025.63k     7123.48k     7172.20k     7147.76k
cast cbc         6000.43k     6978.88k     7123.48k     7147.76k     7123.48k

blowfish cbc [*] 6404.43k     7304.13k     7508.36k     7627.94k     7477.61k
cast cbc    [**] 4404.82k     4909.89k     5000.44k     4993.28k     5013.06k

[*]  pentium pro specific version
[**] pentium specific version

Work still to be done.
- Test vectors for various modes of CAST5 need to be put into casttest.c
- More work on C code variants of the CAST inner loop.
- More testing of the various assember implementations.
- General code cleanups

eric 15-Nov-1997

08-Jan-1998
 - fixes to md5 and sha1 for bignendian machines.