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

Need help on Athlon optimized gemm kernel!



Hello all,

I am currently working on a new Athlon optimized gemm kernel, but I ran 
into a problem:
The kernel crunches 6 dotproducts simultaneously. Of course I must store 
the six produces
elements of matrix(block) C and load the next elements on the stack 
after that operation. And this
is exactly the problem. As long as I leave out the exchange of elements 
of C (means that
the results of all dotproducts in the matrix multiplication are 
accumulated in only 6 stack registers),
the matrix multiplication runs with a stellar speed of 1.93 FLOPS/cycle 
on my Athlon 600 classic/Win2000. When
I insert the exchange part (I have tried some dozens variations of this) 
performance dops enormously,
which I cannot explain.
Currently I try to modify the routine for MSVC++ in order to run AMD 
Code Analyzer, but I do not
think that this will enlight the problem.
So I ask everybody who feels able to help me, to progam the register 
exchange part of
my kernel. I have prepared a NASM .asm file (and C test program) that is 
ready for modifications. This code is the one
that runs with 1.93 FLOPS/cycle so a direct comparion is possible.
Requirements:
- Cygwin installed
- NASM installed
- Skills in Assembly

The person that finds a fast solution will win a golden cake and much 
honor!
If anybody from AMD reads this posting, please help us. Frank S., what's 
about you?

Regards

Julian

R Clint Whaley wrote:

>Guys,
>
>I include below some timings on a 733 Mhz G4e (access courtesy of SourceForge
>compile farm).  For quite a while now, Apple's "half the Mhz, half again the
>price" strategy has eluded me, but this machine ought to at least reduce the
>screaming fits of it's laugh-test failure to at most a few furtive chuckles.
>
>Essentially, it is still not going heads up against either the Athlon or
>P4 (and if anyone hits me with the clock-for-clock crap, I will point out that
>clock for clock the original Power chip is still the champ), but I think
>it is cleaning the floor with the PIII, for instance (let's not mention 
>price, though, eh?).
>
>In single precision, its results are roughly 75% of a P4 clocked at twice
>its speed (before you sneer with the "easy to be fast at low Mhz", I'll remind
>you it is doing this with good ol' SDRAM, so that's pretty impressive), and it
>almost doubles the performance of a 933Mhz PIII . . .
>
>These results are much crappier on an original G4.  Obviously, the extra level
>of cache can't be hurting, but perhaps the greater instruction bandwidth,
>etc., are helping as well.
>
>I found it interesting to compare these timings to the ones I have previously
>posted for the P4 and PIII.  Note that gemm timings can be compared pretty
>directly (no real change from 3.3.0 till 3.3.7), but the LU timings cannot
>(3.3.7 has some speedups over 3.3.0) . . .
>
>Cheers,
>Clint
>
>ATLAS 3.3.7 on 733Mhz G4e, 256K L2, 1MB L3
>
>             100    200    300    400    500    600    700    800    900   1000
>          ====== ====== ====== ====== ====== ====== ====== ====== ====== ======
>ATL dLU    386.8  480.7  513.0  580.7  594.3  684.9  671.8  668.7  703.8  724.1
>ATL dMM    416.7  687.7  771.4  914.3  757.6  919.1  879.5  922.5  928.7  943.4
>
>ATL sLU    437.3  631.0  897.8  982.8 1109.4 1307.5 1343.7 1482.7 1566.4 1586.1
>ATL sMM   1428.6 1600.0 1800.0 2560.0 2500.0 2400.0 2450.0 3011.8 2803.8 2631.6
>
>            1200   1400   1600   1800   2000   2200   2400   2600   2800   3000
>          ====== ====== ====== ====== ====== ====== ====== ====== ====== ======
>ATL dLU    733.3  758.7  786.6  799.7  809.0  819.4  833.0  838.5  840.4  837.0
>
>ATL sLU   1744.4 1846.8 1922.1 1993.0 2058.4 2118.3 2167.8 2206.0 2261.3 2275.0
>ATL sMM   2953.8 2814.4 3022.9 2858.8 3053.4 2937.4 3061.8 2936.7 3081.0 2995.0
>
>_______________________________________________
>Math-atlas-results mailing list
>Math-atlas-results@lists.sourceforge.net
>http://lists.sourceforge.net/lists/listinfo/math-atlas-results
>