Date: Mon, 5 Dec 2011 14:13:38 -0600 From: "jfoug" <jfoug@....net> To: <john-dev@...ts.openwall.com> Subject: RE: 2 changes to dynamic format >On my machine, these changes made dynamic_fmt as fast as the "thick" >experimental raw-md5 (benchmark still 3% slower but real world speed is >same) so I think my "mission" is completed and we could ditch >raw-md5thick again, or move it to unused/. I think we keep the raw-md5thick for now. It is still quite a bit faster on my 32 bit builds, so is a decent test bed. Also, I have used it some to find some optimizations, such as merging mmx-md5.S and intrisics.S code, so it is the same (inside the formats). We can for the 'official' release, put this file into unused, and if people want to use it for testing, they simply need to copy that file into src (I would be one doing that for 'now'). >Another stunt double is "NT2". This format is very interesting for >benchmarking MD4 intrinsics against the old NT format, but it currently >is a little slower on most builds (a lot slower on some). The >problem is it will load "before" the normal NT format because of >filename sort order. So maybe that one too should go to unused/ to avoid >some end-user confusion. Another possibility is to rename it to >something like "x-nt_fmt_plug.c" (x for experimental) so it will come >last in auto load order. I like the x-*_fmt_plug.c naming, as long as it does not clash with any algorithm name. You are not guaranteed to have this format 'last', but it will be pretty darn close to the bottom of the plugin format 'group'. Jim.
Powered by blists - more mailing lists
Powered by Openwall GNU/*/Linux - Powered by OpenVZ