Date: Sat, 7 Feb 1998 23:37:37 -0600 (CST) From: Damon Permezel <dap@damon.com> To: nate@mt.sri.com (Nate Williams) Cc: dap@damon.com, hackers@FreeBSD.ORG Subject: Re: MMX, pentium, etc Message-ID: <199802080537.XAA10571@damon.com> In-Reply-To: <199802080517.WAA17659@mt.sri.com> from Nate Williams at "Feb 7, 98 10:17:33 pm"
next in thread | previous in thread | raw e-mail | index | archive | help
"Nate Williams sez: " > > I have a Java app which uses some intensive FP, for which I have > > written native (aka: C) routines to accelerate. > > > > I am getting to the point where I am once again tapped out, so I want > > to explore hand coded MMX instructions in my (unrolled, hand optimized) > > inner loops. > > Umm, if you're using the FPU, then MMX would be a lose if you mix FPU > and MMX instructions. Better to stick with just FPU instructions. Well, I'll make that decision when I know what the MMX instructions are. All I know is, MMX are supposed to be for multimedia, and in particular enable full-screen s/w mpeg. A lot of what I do is inverse DCT, except I go 256x256, rather than 8x8. I will probably change to an integer iDCT when I am able to see what the MMX instructions are: ie if they are all integer instructions. I suppose I should change to integer anyway, but I have been spoiled by working on machines which had decent FP performance. To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199802080537.XAA10571>