[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Altivec matmul kernel (attachment)


1 - no complex support (haven't had the time).
2 - altivec doesn't support double precision.
2.5 - not sure what you mean here.
3 - yes, I've successfully built the full lib.  With the Altivec "Java 
mode" turned off (see another post of mine to atlas-comm for info on 
this) on my 533 Mhz G4 machine, I get about 1870 Mflops peak sgemm.
4 - I believe that gcc 3.0 has Altivec support also, although I could be 


On Thursday, July 19, 2001, at 03:36  PM, Camm Maguire wrote:

> Greetings, and thanks for your reply.  That works!  A few questions:
> 1) Any complex support yet?
> 2) Does altivec have any double precision support, a la SSE2
> 2.5) Any l2/l1 code?
> 3) Have you successfully build the full lib using this?
> 4) The gcc/as situation is indeed bad.  I've confirmed that the latest
> standard binutils can't assemble these instructions yet.  Therefore, to
> build on Debian, at least the patched binutils needs to be packaged
> and accepted as an option for powerpc before atlas can be built there
> with this kernel.
> Take care,
> Nicholas Coult <coult@augsburg.edu> writes:
>> Hi,
>> Sorry I'm having some email troubles just now.  But I see your message
>> regarding the patched gcc in the developer mail archive.   The problem
>> is that you need to add the flag '-fvec' to the gcc command-line flags.
>> This enables the altivec extensions in gcc.  (I might be wrong about
>> -fvec; it might be -fvector or -faltivec, but in any case you need to
>> explicitly turn on the altivec features).
>> -Nick
Nicholas Coult, Ph.D.,  web: http://melby.augsburg.edu/~coult
Assistant Professor, Department of Mathematics, Augsburg College
coult@augsburg.edu, phone:  (612) 330-1064 office: Science Hall 137B
> --
> Camm Maguire			     			camm@enhanced.com
> =========================================================================
> =
> "The earth is but one country, and mankind its citizens."  --  
> Baha'u'llah