sha/asm/sha*-mb-x86_64.pl: commentary update.

diff --git a/crypto/sha/asm/sha1-mb-x86_64.pl b/crypto/sha/asm/sha1-mb-x86_64.pl
index fff96547da4bdd0c9d7a8e2b7d9b047419102375..93996e15f96e93746c32e23170f8409cbbe0b96c 100644 (file)
--- a/crypto/sha/asm/sha1-mb-x86_64.pl
+++ b/crypto/sha/asm/sha1-mb-x86_64.pl
@@ -12,20 +12,22 @@
 # naturally limited to 4 on pre-AVX2 processors and to 8 on
 # AVX2-capable processors such as Haswell.
 #
-#              this    +aesni(*)       sha1    aesni-sha1      gain
+#              this    +aesni(i)       sha1    aesni-sha1      gain(iv)
 # -------------------------------------------------------------------
-# Westmere(**) 10.4/n  +1.28=3.88(n=4) 5.44    6.58            +70%
-# Atom(**)     18.9/n  +3.93=8.66(n=4) 10.0    14.0            +62%
+# Westmere(ii) 10.4/n  +1.28=3.88(n=4) 5.44    6.58            +70%
+# Atom(ii)     18.9/n  +3.93=8.66(n=4) 10.0    14.0            +62%
 # Sandy Bridge (8.16   +5.15=13.3)/n   4.99    5.98            +80%
 # Ivy Bridge   (8.03   +5.14=13.2)/n   4.60    5.54            +68%
-# Haswell(***) (8.96   +5.00=14.0)/n   3.57    4.55            +160%
+# Haswell(iii) (8.96   +5.00=14.0)/n   3.57    4.55            +160%
 # Bulldozer    (9.75   +5.76=15.5)/n   5.95    6.37            +64%
 #
-# (*)  multi-block CBC encrypt with 128-bit key;
-# (**) (HASH+AES)/n does not apply to Westmere for n>3 and Atom,
+# (i)  multi-block CBC encrypt with 128-bit key;
+# (ii) (HASH+AES)/n does not apply to Westmere for n>3 and Atom,
 #      because of lower AES-NI instruction throughput;
-# (***)        "this" is for n=8, when we gather twice as much data, result
+# (iii)        "this" is for n=8, when we gather twice as much data, result
 #      for n=4 is 7.98+4.44=12.4;
+# (iv) improvement coefficients in real-life application are somewhat
+#      lower and range from 30% to 100% (on Haswell);
 
 $flavour = shift;
 $output  = shift;

diff --git a/crypto/sha/asm/sha256-mb-x86_64.pl b/crypto/sha/asm/sha256-mb-x86_64.pl
index 5bed43b60cac292e1c0e9fee778d1b4c2ffd9f55..2e4b102f52829a41ff6811ad30bb4eadd3c224ed 100644 (file)
--- a/crypto/sha/asm/sha256-mb-x86_64.pl
+++ b/crypto/sha/asm/sha256-mb-x86_64.pl
@@ -12,21 +12,23 @@
 # naturally limited to 4 on pre-AVX2 processors and to 8 on
 # AVX2-capable processors such as Haswell.
 #
-#              this    +aesni(*)       sha256  aesni-sha256    gain
+#              this    +aesni(i)       sha256  aesni-sha256    gain(iv)
 # -------------------------------------------------------------------
-# Westmere(**) 23.3/n  +1.28=7.11(n=4) 12.3    +3.75=16.1      +126%
-# Atom(**)     39.1/n  +3.93=13.7(n=4) 20.8    +5.69=26.5      +93%
+# Westmere(ii) 23.3/n  +1.28=7.11(n=4) 12.3    +3.75=16.1      +126%
+# Atom(ii)     39.1/n  +3.93=13.7(n=4) 20.8    +5.69=26.5      +93%
 # Sandy Bridge (20.5   +5.15=25.7)/n   11.6    13.0            +103%
 # Ivy Bridge   (20.4   +5.14=25.5)/n   10.3    11.6            +82%
-# Haswell(***) (21.0   +5.00=26.0)/n   7.80    8.79            +170%
+# Haswell(iii) (21.0   +5.00=26.0)/n   7.80    8.79            +170%
 # Bulldozer    (21.6   +5.76=27.4)/n   13.6    13.7            +100%
 #
-# (*)  multi-block CBC encrypt with 128-bit key;
-# (**) (HASH+AES)/n does not apply to Westmere for n>3 and Atom,
+# (i)  multi-block CBC encrypt with 128-bit key;
+# (ii) (HASH+AES)/n does not apply to Westmere for n>3 and Atom,
 #      because of lower AES-NI instruction throughput, nor is there
 #      AES-NI-SHA256 stitch for these processors;
-# (***)        "this" is for n=8, when we gather twice as much data, result
+# (iii)        "this" is for n=8, when we gather twice as much data, result
 #      for n=4 is 20.3+4.44=24.7;
+# (iv) improvement coefficients in real-life application are somewhat
+#      lower and range from 75% to 130% (on Haswell);
 
 $flavour = shift;
 $output  = shift;