[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 25 Apr 2013 15:40:58 -0500
From: "jfoug" <jfoug@....net>
To: <john-dev@...ts.openwall.com>
Subject: RE: ICC performance regression
Here are my timings. Cygwin 32 bit, i7-2600 @ 3.40GHz:
New ICC sse2 md5: 30525 dynamic_0 27042K (Built on my VM, not the files
you built).
Old ICC sse2 md5: 30685 dynamic_0 27191K
gcc 4.7.2 x64 32-bit cross compile sse2 md5: 31575 dynamic_0 26949K
Here are my timings on ubuntu 12.10 (running VM, no OMP builds):
New ICC sse2 md5: 38081 dynamic_0 30906K
Old ICC sse2 md5: N/A dynamic_0 29974K
gcc 4.7.2 x64 32-bit cross compile sse2 md5: 34950 dynamic_0 27720K
The gcc build is much bigger. But having a working icc environment, I will
look at carrying forward. Now that I can build and test, I will look at
some of the changes we had talked about offline. Providing a 'usual'
interlaced input/output interface. Providing a flat 'scalar' interface.
Possibly even providing a multi-input CTX like interface. However, as we
have seen from experience, it usually works out that a huge amount of gain,
comes from the calling format, doing a fast job of input/output buffer
handling, and letting the crypt code, just perform the crypt on 1 (para)
block of prepared data.
But I can do a lot better, having an environment myself, for doing the
builds. Last time I looked at making any mods, when we reduced the temp
buffers in SHA1, I did not have a icc (or current linux x64) build
environment. I have those now, so can do a lot more playing around with
that file, and now work to get sha2 functions added also.
Jim.
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
