x86[_64] assembly pack: update benchmark results.

#		AES-128-CBC	+SHA1		stitch      gain

# Westmere	3.77[+5.6]	9.37		6.65	    +41%

# Sandy Bridge	5.05[+5.2(6.3)]	10.25(11.35)	6.16(7.08)  +67%(+60%)

# Ivy Bridge	5.05[+4.7]	9.75		5.59        +74%

# Bulldozer	5.77[+6.1]	11.87		6.47        +83%

#

#		AES-192-CBC

# Westmere	4.51		10.11		6.97	    +45%

# Sandy Bridge	6.05		11.25(12.35)	6.34(7.27)  +77%(+70%)

# Ivy Bridge	6.05		10.75		6.07        +77%

# Bulldozer	6.89		12.99		7.02        +85%

#

#		AES-256-CBC

# Westmere	5.25		10.85		7.25	    +50%

# Sandy Bridge	7.05		12.25(13.35)	7.06(7.70)  +74%(+73%)

# Ivy Bridge	7.05		11.75		7.12        +65%

# Bulldozer	8.00		14.10		8.24        +71%

#

# (*)	There are two code paths: SSSE3 and AVX. See sha1-568.pl for

#	background information. Above numbers in parentheses are SSSE3

#		AES-128-CBC	AES-192-CBC	AES-256-CBC

# Westmere	1.31		1.55		1.80

# Sandy Bridge	0.93		1.06		1.22

# Ivy Bridge	0.92		1.06		1.21

# Bulldozer	0.76		0.90		1.04

$flavour = shift;

$output  = shift;

# one byte out of 8KB with 128-bit key, Sandy Bridge - 0.97. Just like

# in CTR mode AES instruction interleave factor was chosen to be 6x.

######################################################################

# For reference, AMD Bulldozer spends 5.77 cycles per byte processed

# with 128-bit key in CBC encrypt and 0.76 cycles in CBC decrypt, 0.70

# in ECB, 0.94 in CTR, 0.95 in XTS... This means that aes[enc|dec]

# instruction latency is 9 cycles and that they can be issued every

# cycle.

$PREFIX="aesni";	# if $PREFIX is set to "AES", the script

			# generates drop-in replacement for

			# crypto/aes/asm/aes-x86_64.pl:-)

# P4		125/125		17.8		84(***)

# Opteron	66 /70		10.1		30

# Core2		54 /67		8.4		18

# Atom		105/105		16.8		53

# VIA Nano	69 /71		13.0		27

#

# (*)	gcc 3.4.x was observed to generate few percent slower code,

#	which is one of reasons why 2.95.3 results were chosen,

# similar manner resulted in almost 20% degradation on Sandy Bridge,

# where original 64-bit code processes one byte in 1.95 cycles.

#####################################################################

# For reference, AMD Bulldozer processes one byte in 1.98 cycles in

# 32-bit mode and 1.89 in 64-bit.

$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;

push(@INC,"${dir}","${dir}../../perlasm");

require "x86asm.pl";

# P4		28.6		14.0		+100%

# Opteron	19.3		7.7		+150%

# Core2		17.8		8.1(**)		+120%

# Atom		31.6		16.8		+88%

# VIA Nano	21.8		10.1		+115%

#

# (*)	comparison is not completely fair, because C results are

#	for vanilla "256B" implementation, while assembler results

# Westmere	5.1/+94%(**)

# Sandy Bridge	5.0/+8%

# Atom		12.6/+6%

# VIA Nano	6.4/+9%

# Ivy Bridge	4.9/0%

# Bulldozer	4.9/+15%

#

# (*)	PIII can actually deliver 6.6 cycles per byte with MMX code,

#	but this specific code performs poorly on Core2. And vice