Commit b7f5503f authored by Andy Polyakov's avatar Andy Polyakov
Browse files

Skylake performance results. 



Reviewed-by: default avatarMatt Caswell <matt@openssl.org>
parent c9c84a13

crypto/aes/asm/aesni-sha1-x86_64.pl

+2 −0
Original line number Diff line number Diff line
@@ -25,6 +25,7 @@ 
# Sandy Bridge	5.05[+5.0(6.1)]	10.06(11.15)	5.98(7.05)  +68%(+58%)

# Ivy Bridge	5.05[+4.6]	9.65		5.54        +74%

# Haswell	4.43[+3.6(4.2)]	8.00(8.58)	4.55(5.21)  +75%(+65%)

# Skylake	2.63[+3.5(4.1)]	6.17(6.69)	4.23(4.44)  +46%(+51%)

# Bulldozer	5.77[+6.0]	11.72		6.37        +84%

#

#		AES-192-CBC
@@ -39,6 +40,7 @@ 
# Sandy Bridge	7.05		12.06(13.15)	7.12(7.72)  +69%(+70%)

# Ivy Bridge	7.05		11.65		7.12        +64%

# Haswell	6.19		9.76(10.34)	6.21(6.25)  +57%(+65%)

# Skylake	3.62		7.16(7.68)	4.56(4.76)  +57%(+61$)

# Bulldozer	8.00		13.95		8.25        +69%

#

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

crypto/aes/asm/aesni-sha256-x86_64.pl

+1 −0
Original line number Diff line number Diff line
@@ -25,6 +25,7 @@ 
# Sandy Bridge	    5.05/6.05/7.05+11.6		13.0	+28%/36%/43%

# Ivy Bridge	    5.05/6.05/7.05+10.3		11.6	+32%/41%/50%

# Haswell	    4.43/5.29/6.19+7.80		8.79	+39%/49%/59%

# Skylake	    2.62/3.14/3.62+7.70		8.10	+27%/34%/40%

# Bulldozer	    5.77/6.89/8.00+13.7		13.7	+42%/50%/58%

#

# (*)	there are XOP, AVX1 and AVX2 code pathes, meaning that

crypto/aes/asm/aesni-x86_64.pl

+1 −0
Original line number Diff line number Diff line
@@ -165,6 +165,7 @@ 
# Westmere	3.77/1.25	1.25	1.25	1.26

# * Bridge	5.07/0.74	0.75	0.90	0.85

# Haswell	4.44/0.63	0.63	0.73	0.63

# Skylake	2.62/0.63	0.63	0.63	0.63

# Silvermont	5.75/3.54	3.56	4.12	3.87(*)

# Bulldozer	5.77/0.70	0.72	0.90	0.70

#

crypto/modes/asm/aesni-gcm-x86_64.pl

+5 −4
Original line number Diff line number Diff line
@@ -22,10 +22,11 @@ 
# [1] and [2], with MOVBE twist suggested by Ilya Albrekht and Max

# Locktyukhin of Intel Corp. who verified that it reduces shuffles

# pressure with notable relative improvement, achieving 1.0 cycle per

# byte processed with 128-bit key on Haswell processor, and 0.74 -

# on Broadwell. [Mentioned results are raw profiled measurements for

# favourable packet size, one divisible by 96. Applications using the

# EVP interface will observe a few percent worse performance.]

# byte processed with 128-bit key on Haswell processor, 0.74 - on

# Broadwell, 0.63 - on Skylake... [Mentioned results are raw profiled

# measurements for favourable packet size, one divisible by 96.

# Applications using the EVP interface will observe a few percent

# worse performance.]

#

# [1] http://rt.openssl.org/Ticket/Display.html?id=2900&user=guest&pass=guest

# [2] http://www.intel.com/content/dam/www/public/us/en/documents/software-support/enabling-high-performance-gcm.pdf

crypto/modes/asm/ghash-x86_64.pl

+3 −2
Original line number Diff line number Diff line
@@ -64,6 +64,7 @@ 
# Ivy Bridge	1.80(+7%)

# Haswell	0.55(+93%) (if system doesn't support AVX)

# Broadwell	0.45(+110%)(if system doesn't support AVX)

# Skylake	0.44(+110%)(if system doesn't support AVX)

# Bulldozer	1.49(+27%)

# Silvermont	2.88(+13%)
@@ -74,8 +75,8 @@ 
# CPUs such as Sandy and Ivy Bridge can execute it, the code performs

# sub-optimally in comparison to above mentioned version. But thanks

# to Ilya Albrekht and Max Locktyukhin of Intel Corp. we knew that

# it performs in 0.41 cycles per byte on Haswell processor, and in

# 0.29 on Broadwell.

# it performs in 0.41 cycles per byte on Haswell processor, in

# 0.29 on Broadwell, and in 0.36 on Skylake.

#

# [1] http://rt.openssl.org/Ticket/Display.html?id=2900&user=guest&pass=guest