diff options
author | Andy Polyakov <appro@openssl.org> | 2005-07-14 13:16:31 +0000 |
---|---|---|
committer | Andy Polyakov <appro@openssl.org> | 2005-07-14 13:16:31 +0000 |
commit | b4f5e5c95912e753147c2efe2f5eb7be97908556 (patch) | |
tree | 74f889e45d0313f1354e75f9535545a3ef5d50a6 /crypto/aes/asm | |
parent | 0d2848b3bafc7473b26597764ed5c65fb42e2a45 (diff) |
Commentary section update.
Diffstat (limited to 'crypto/aes/asm')
-rwxr-xr-x | crypto/aes/asm/aes-x86_64.pl | 19 |
1 files changed, 14 insertions, 5 deletions
diff --git a/crypto/aes/asm/aes-x86_64.pl b/crypto/aes/asm/aes-x86_64.pl index 3113e35dc6..f39a36ab3d 100755 --- a/crypto/aes/asm/aes-x86_64.pl +++ b/crypto/aes/asm/aes-x86_64.pl @@ -10,12 +10,21 @@ # # aes-*-cbc benchmarks are improved by 50% [compared to gcc 3.3.2 on # Opteron 240 CPU] plus all the bells-n-whistles from 32-bit version -# [you'll notice a lot of resemblance], such as compressed S-boxs -# and prefetch of these tables in CBC mode, as well as avoiding L1 -# cache aliasing between stack frame and key schedule and already -# mentioned tables. +# [you'll notice a lot of resemblance], such as compressed S-boxes +# in little-endian byte order, prefetch of these tables in CBC mode, +# as well as avoiding L1 cache aliasing between stack frame and key +# schedule and already mentioned tables. # -# ECB performance is 15.6 cycles per processed byte for 128-bit key. +# Performance in number of cycles per processed byte for 128-bit key: +# +# ECB CBC encrypt +# AMD64 15.6 14.6(*) +# EM64T 23.3(**) 21.4(*) +# +# (*) CBC benchmarks are better than ECB thanks to custom ABI used +# by the private block encryption function. +# (**) This module exhibits virtually same ECB performance as 32-bit +# counterpart on [current] Intel CPU. $output=shift; open STDOUT,"| $^X ../perlasm/x86_64-xlate.pl $output"; |