On Tuesday 09 December 2003 08:28, Andy Polyakov wrote: > This is a poll for votes. > > It was noted that [at least] Intel IA-32 compiler, linux-ia32-icc > target, generates *noticeably*, 30% to be specific, faster code for SHA1 > than hand-coded assembler implementation on at least P4 platform. I have > re-tuned SHA1 assembler implementation which now performs as following: > > compared with current compared with icc > assembler impl. generated code > Pentium -25% +37% > PIII/AMD +8% +16% > P4 +85%(!) +45% > > Options for integrating re-tuned code are: > > 1. replace crypto/sha/asm/sha1-586.pl and let couple of Pentium users > suffer 25% performance loss; > 2. add crypto/sha/asm/sha1-686.pl, make it default, so that couple of > Pentium users *can* pull old code if they need 25% back; > 3. add crypto/sha/asm/sha1-686.pl and have ./config choose between two > versions, depending on which computer ./config is executed; > > My personal vote is #1. If nobody speaks up within 3-4 days, I'll > replace sha1-586.pl with another version. A.
How big a hassle is it to do #2? True, the performance enhancement on the newer, faster machines is pretty impressive. However, it's the older, slower machines that are most in need of speedups. If support for the older hardware can be maintained without a huge headache, I think that's the correct path. ______________________________________________________________________ OpenSSL Project http://www.openssl.org Development Mailing List [EMAIL PROTECTED] Automated List Manager [EMAIL PROTECTED]
