Athlon optimized BLAS routines
From: Emil Briggs <briggs \\at// tick.physics.ncsu.edu>
I've written a set of Athlon optimized double precision
BLAS1 routines in assembly. The source can be found at
http://nemo.physics.ncsu.edu/~briggs/blas_src_v0.11_tar.gz
The routines are in an early stage of development but they do
pass the Level 1 Blas testing routines from netlib and they
work correctly in the applications I've tested. Thats not
a guarantee that they are correct in all cases however so
use with caution. The routines use the Athlon memory prefetch
instructions which really boost performance for large data
sets. (You'll get an illegal instruction exception on PentiumIII
systems).
I'm still tuning, optimizing and debugging but these are already
much faster on an Athlon than any other set of BLAS1 libs that
I know of available for Linux.
Regards
Emil
------------------------------------------------------------------
Benchmark results on a 500Mhz Athlon system. Best time of 20 iterations.
Intel optimized libraries are version lsblasppro1.1o_08.99.a courtesy
of Greg Henry.
DAXPY
**********************************************************
N Athlon libs Intel optimized
4096 411.49 373.48
16384 109.60 82.35
65536 57.49 49.29
262144 52.66 43.67
1048576 51.31 43.66
DDOT
**********************************************************
N Athlon libs Intel optimized
4096 587.35 164.40
16384 155.30 72.18
65536 80.81 42.39
262144 75.72 28.76
1048576 74.79 26.16
DSCAL
**********************************************************
N Athlon libs Intel optimized
4096 293.67 164.40
16384 91.50 72.18
65536 58.36 42.39
262144 44.73 28.76
1048576 42.25 26.16
DNRM2
**********************************************************
N Athlon libs Intel optimized
4096 630.45 545.39
16384 394.94 338.10
65536 182.55 116.71
262144 156.92 72.56
1048576 153.80 65.74
DCOPY
**********************************************************
N Athlon libs Intel optimized
4096 4398.05 2421.83
16384 888.85 515.11
65536 539.97 324.83
262144 449.60 308.21
1048576 445.05 298.25
-------------------------------------------------------------------
To unsubscribe send a message body containing "unsubscribe"
to beowulf-request \\at// beowulf.org