Re: SSE Level 3 drop in gemm

[Camm Maguire <camm@enhanced.com>]

Hello again!  Sorry to write so frequently, but this is almost wrapped
up, and I'd like to get it out quickly if possible, while its fresh in
my mind.

R Clint Whaley <rwhaley@cs.utk.edu> writes:

> Camm,
> 
> >Now I have a different issue.  My kernel likes nb=56 the best.  Atlas
> >standard likes nb=64.  And this is what I get in sMMRES:
> >
> >intech20:/mnt/i19/f/debian/mm/atlas/tmp/atlas-3.1.2D/tune/blas/gemm/Linux_fpic$ cat res/sMMRES
> >MULADD  LAT  NB  MU  NU  KU  FFTCH  IFTCH  NFTCH    MFLOP
> >     0    2  64   5   1  64      0      5      1   371.46
> >16
> >ATL_sgemm_SSE.c "CM"
> >     1    1  64   2   2  64      0      4      1   617.61
> 
> The best solution is probably to add another case to scases.dsc.  I guess
> you have a line that looks something like this right now:
> 0 1 1 1 1 1 2 2 64 ATL_sgemm_SSE.c "CM"
> add a second line, with the explicit 56 blocking factor:
> 0 -56 -56 -56 1 1 2 2 64 ATL_sgemm_SSE.c "CM"
> 

OK.  Thanks!  But even with the sMMRES above, doing the make install
made a lib without my kernel.  bin/arch/xsl3blastst gives 370 MFLOPS. :-(

> And that should force ATLAS time using your best blocking factor, which should
> then be selected . . .  In the long run, I think it makes sense to have ATLAS
> do this: if a user-contributed kernel is found to beat the generated kernel,
> run through the list of possible NBs again, to find the best.  I'll look at
> adding that while I'm doing the cleanup stuff . . .
> 

Sounds good!

> >Also, any sugegstions on the unrolling issue?  If I put in macros to
> >unroll k at differing levels depending on KB, would that confuse the
> >search engine?  Should install faster than having 3 different k unroll
> >kernels to time.
> 
> I don't think it would confuse the script, but you'd have to force the
> different NB's at present via a technique like above.  Should work like
> a charm if I add the NB search, though . . .
> 

OK.

> Cheers,
> Clint
> 
> 

Just a reminder, one other outstanding issue is the data alignment.
Setting Atlas_Cachelen did not seem to affect the mmtst.c or fc.c
programs.  Am I missing something here?

Take care,


-- 
Camm Maguire			     			camm@enhanced.com
==========================================================================
"The earth is but one country, and mankind its citizens."  --  Baha'u'llah