| speed comparison of IDL, numPy, Matlab [message #23666] |
Mon, 05 February 2001 10:18  |
Benyang Tang
Messages: 3
Registered: May 2000
|
Junior Member |
|
|
Out of curiosity, I did a quick benchmark test of IDL, NumPy and Matlab on my
desktop machine. I know benchmarking is a complicated issue; don't take my
naive test too serious.
======================
Conditions of the test:
======================
* Machine: a dual Intel Xeon 550 MHz box with 1GB ram, running RedHat Linux
6.2. The machine was not doing any serious service, so the test code should
have had close to 100% of the resources.
* IDL, version 5.3
* Python version 1.5.2; NumPy release 15.3-1
* Matlab version 5.3
* For each code, I ran it several times so the timing became somewhat stable. I
just took the last reading; for the size of 600X600, timing fluctuation of
less than 2% was observed.
* All tests were done in a period of 20 minutes; within this period the
following data were collected. About 2 hours later, I redid the tests and
obtained similar results.
============================================================ =====
Timing (in seconds) of matrix multiplication of 2 m-by-m matrixes
============================================================ =====
m 100 200 300 400 500 600
-----------------------------------------
single precision
IDL 0.02 0.12 0.54 1.83 4.06 6.90
NumPy 0.01 0.12 0.46 1.37 3.05 5.49
Matlab (1)
double precision
IDL 0.04 0.41 1.89 5.14 10.72 18.61
NumPy 0.02 0.16 0.97 2.50 5.05 8.75
Matlab 0.01 0.15 0.94 2.50 4.90 8.69
-----------------------------------------
m 100 200 300 400 500 600
(1) Matlab does not do single-precision calculation.
===================
What the tests tell:
===================
1) NumPy is about as fast as Matlab;
2) NumPy is about 25% faster (in single precision), or more than 100% faster
(in double precision) than IDL.
==================================================
Here are the codes (for the double precision test):
==================================================
IDL:
===
for m = 100,600,100 do begin
a = double(randomn(0,m,m))
b = double(randomn(0,m,m))
time0 = systime(1)
c = a##b
dtime = systime(1)-time0
print, string(m,m,dtime, format='("multiplication of " ,i3, "X", i3, " matrixes takes ", f5.2)')
endfor
NumPy:
=====
from Numeric import *
from RandomArray import *
import time
for m in [100,200,300,400,500,600]:
a= uniform(-1,1,(m,m,))
b= uniform(-1,1,(m,m,))
a=a.astype(Float64)
b=b.astype(Float64)
time0 = time.time()
c = matrixmultiply(a,b)
dtime = time.time() - time0
print 'multiplication of %dX%d matrixes takes %5.2f' %(m,m,dtime)
Matlab:
======
for m = [100 200 300 400 500 600];
a = randn(m);
b = randn(m);
time0 = clock;
c = a*b;
dtime = etime(clock,time0);
fprintf('multiplication of %dX%d matrixes takes %5.2f\n', m,m,dtime)
end
--
<> Benyang Tang
<> 300-323, JPL
<> 4800 Oak Grove Drive
<> Pasadena, CA 91109, USA
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
