| x*x versus x^2 [message #61206] |
Wed, 09 July 2008 09:32  |
Conor
Messages: 138
Registered: February 2007
|
Senior Member |
|
|
So I've been looking at execution time for various algorithms, and I
found this interesting result:
bigarr = fltarr(1000,1000)
t1 = systime(/seconds)
t = bigarr^2.0
t2 = systime(/seconds)
t = bigarr*bigarr
t3 = systime(/seconds)
print,t2-t1
print,t3-t2
IDL prints:
0.024163008
0.010262012
Apparently multiplying an array by itself is twice as fast as using
the carat operator! Anyone know why this is? Is it a memory issue or
something?
|
|
|
|
| Re: x*x versus x^2 [message #61285 is a reply to message #61206] |
Sat, 12 July 2008 12:30  |
Sven Geier
Messages: 17
Registered: July 2002
|
Junior Member |
|
|
Conor wrote:
> So I've been looking at execution time for various algorithms, and I
> found this interesting result:
>
> bigarr = fltarr(1000,1000)
>
> t1 = systime(/seconds)
> t = bigarr^2.0
> t2 = systime(/seconds)
> t = bigarr*bigarr
> t3 = systime(/seconds)
>
> print,t2-t1
> print,t3-t2
>
>
> IDL prints:
>
> 0.024163008
> 0.010262012
>
> Apparently multiplying an array by itself is twice as fast as using
> the carat operator! Anyone know why this is? Is it a memory issue or
> something?
What version of IDL and what OS are we talking about here? I tried to
reproduce this (IDL 6.3 Win32/x86) and I get wildly varying results every
time I hit F5:
IDL> .GO
0.016000032
0.014999866
IDL> .GO
0.00000000
0.016000032
IDL> .GO
0.016000032
0.00000000
The help on SYSTIME() tels me that on windows this simply doesn't seem to
have the accuracy to measure quantities in the small-number-of-milliseconds
range...
When I increase the array sizes by a factor of 10 in each dimension I get
the following:
IDL> .GO
0.45400000
0.90599990
IDL> .GO
0.62500000
0.75000000
IDL> .GO
0.64000010
0.76600003
IDL> .GO
0.62500000
0.75000000
Not exactly repeatable.
However this is all based on multiplication of zero with zero -- to what
degree is that optimized under the hood? If I replace the first line in
your script
bigarr = fltarr(1000,1000)
with
bigarr = randomn(seed,10000,10000)
I get a completely different picture:
IDL> .GO
5.1099999
0.75000000
IDL> .GO
5.2340000
0.76600003
IDL> .GO
5.1559999
0.78100014
IDL> .GO
5.1410000
0.76600003
Replacing the float exponent "2.0" with a plain "2" turns this into:
IDL> .GO
0.65700006
0.75000000
IDL> .GO
0.64100003
0.75000000
IDL> .GO
0.65700006
0.76500010
IDL> .GO
0.64100003
0.76500010
Finally:
replacing bigarr^2 and bigarr*bigarr
with bigarr*2 and bigarr+bigarr
Gives me the following:
IDL> .GO
0.53200006
0.76500010
IDL> .GO
0.53099990
0.75000000
IDL> .GO
0.54700017
0.75000000
IDL> .GO
0.51500010
0.76599979
IDL> .GO
0.50000000
0.78200006
IDL> .GO
0.48399997
0.76600003
Just to add some data points...
--
http://www.sgeier.net
My real email address does not contain any "Z"s.
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
