Re: BLAS by Kazushige Goto Kazushige Goto

[David Bateman]

If you've compiled with --enable-shared you shouldn't need to recompile
octave to use this library. However I dodn't see much value. With the
attached benchmarking software I tried three cases

case 1) octave 2.1.46 without ALTAS
case 2) octave 2.1.46 with ATLAS
case 3) octave 2.1.46 with the command line 
         LD_PRELOAD=/tmp/libgoto_p3_512-r0.6.so octave -q

The results are below... You'll note that the improvement is basically
15% on my laptop P3 wrt ATLAS. With the fact that this code is only
available as a binary, and that it appears not to have been cleanly
implemented (see undefined symbols below), I consider the 15% gain as
not worth the hassles.

Regards
David

case 1
------
octave:1>  mybench
    OCTAVE/Matlab benchmark version bm 0.1
    Speed of octave 2.1.46 on i686-pc-linux-gnu 
    relative to Octave 2.0.16 on PII 300MHz

Matrix inversion (LINPACK/LAPACK)        4.91 ( 0.21 sec) +/- 1.8% (98 runs)
Schur decomposition (LAPACK)             5.40 ( 0.14 sec) +/- 11.0% (400 runs)
Linear equation (LINPACK/LAPACK)         5.70 ( 0.09 sec) +/- 1.3% (231 runs)
Linear equation N RHS (LINPACK/LAPACK)   5.76 ( 0.09 sec) +/- 2.8% (231 runs)
Matrix determinant (LINPACK/LAPACK)      5.54 ( 0.08 sec) +/- 4.8% (420 runs)
Matrix Multiply (BLAS)                   5.77 ( 0.26 sec) +/- 4.4% (77 runs)
Fourier transforms (FFTPACK)             4.76 ( 0.08 sec) +/- 1.4% (252 runs)
for loop                                 2.51 ( 0.08 sec) +/- 2.3% (245 runs)
-- Performance index (bm 0.1): 4.9


case 2
------
octave:1> mybench
    OCTAVE/Matlab benchmark version bm 0.1
    Speed of octave 2.1.46 on i686-pc-linux-gnu 
    relative to Octave 2.0.16 on PII 300MHz

Matrix inversion (LINPACK/LAPACK)       13.63 ( 0.08 sec) +/- 1.0% (280 runs)
Schur decomposition (LAPACK)             6.02 ( 0.12 sec) +/- 3.1% (175 runs)
Linear equation (LINPACK/LAPACK)        12.56 ( 0.04 sec) +/- 1.4% (560 runs)
Linear equation N RHS (LINPACK/LAPACK)  12.93 ( 0.04 sec) +/- 3.1% (539 runs)
Matrix determinant (LINPACK/LAPACK)     12.31 ( 0.04 sec) +/- 1.0% (539 runs)
Matrix Multiply (BLAS)                  26.24 ( 0.06 sec) +/- 0.5% (371 runs)
Fourier transforms (FFTPACK)             4.68 ( 0.09 sec) +/- 0.8% (245 runs)
for loop                                 2.48 ( 0.08 sec) +/- 3.1% (252 runs)
-- Performance index (bm 0.1): 9.17


case 3
------
sh: relocation error: /tmp/libgoto_p3_512-r0.6.so: undefined symbol: sqrt
warning: broken pipe
octave:1>  mybench
    OCTAVE/Matlab benchmark version bm 0.1
    Speed of octave 2.1.46 on i686-pc-linux-gnu 
    relative to Octave 2.0.16 on PII 300MHz

Matrix inversion (LINPACK/LAPACK)       15.16 ( 0.07 sec) +/- 4.7% (308 runs)
Schur decomposition (LAPACK)             6.43 ( 0.11 sec) +/- 1.1% (189 runs)
Linear equation (LINPACK/LAPACK)        14.17 ( 0.03 sec) +/- 0.9% (588 runs)
Linear equation N RHS (LINPACK/LAPACK)  14.32 ( 0.04 sec) +/- 2.7% (588 runs)
Matrix determinant (LINPACK/LAPACK)     14.24 ( 0.03 sec) +/- 5.2% (1720 runs)
Matrix Multiply (BLAS)                  29.80 ( 0.05 sec) +/- 0.7% (420 runs)
Fourier transforms (FFTPACK)             4.73 ( 0.08 sec) +/- 1.2% (245 runs)
for loop                                 2.47 ( 0.09 sec) +/- 2.3% (238 runs)
-- Performance index (bm 0.1): 9.97


-- 
David Bateman                                address@hidden
Motorola CRM                                 +33 1 69 35 25 00 (Ph)
Espace Technologique, Commune de St Aubin    +33 1 69 35 25 01 (Fax)
91193 Gif-Sur-Yvette FRANCE

The information contained in this communication has been classified as: 

[x] General Business Information 
[ ] Motorola Internal Use Only 
[ ] Motorola Confidential Proprietary

mybench.m
Description: Text document

bm_test.m
Description: Text document