Re: atlas 3.4.1 and AltiVec
On Die, 2002-12-03 at 01:09, Camm Maguire wrote:
> Greetings! It had always been my hope to generate an atlas2-altivec
> package when I got around to updating to 3.4.1. Unfortunately, the
> support in atlas for this is based on the C extension protocol
> (e.g. 'vector'). I notice that gcc-3.2 can (optionally?) generate
> some altivec instructions, but cannot parse this syntax. I have
> downloaded the patched gcc and confirmed that it builds a faster atlas
> passing all tests. My plan was to save the assembler files thus
> compiled, and pass them to the normal gcc as alternatives to be timed
> and included in the library. Of course the easiest way for me would
> be if there was a way to have the standard gcc compile these extension
> files, which would have to work even on an autobuilder. Lastly, I
> could hope that gcc with -maltivec is good enough to beat or basically
> match the hand written altivec C code that comes with atlas, though
> this sounds implausible.
I didn't think -maltivec generated any altivec code at all without using
special vector functions?
> Am I stuck with the first option?
I'm by no means an expert, but that's my understanding. To get in touch
with gcc experts, you may want to post to the linuxppc-dev or even gcc
Earthling Michel Dänzer (MrCooper)/ Debian GNU/Linux (powerpc) developer
XFree86 and DRI project member / CS student, Free Software enthusiast