| Re: Cumulative max() in *arbitrary* dimension? [message #79509 is a reply to message #79426] |
Thu, 08 March 2012 10:33   |
JDS
Messages: 94
Registered: March 2009
|
Member |
|
|
On Tuesday, February 28, 2012 1:57:20 PM UTC-5, Gianguido Cianci wrote:
> On Monday, February 27, 2012 6:09:38 PM UTC-6, JDS wrote:
>> In fact, I was very surprised to find
>> that, when working along the final dimension of large arrays (of a few
>> hundred million elements), MAX_CUMULATIVE is ~2x faster than its
>> MAX(DIMENSION=) analog, which produces a subset of the information!
>
> Mind. Blown.
I've since tuned this up a bit more, saving 1/2 of the index computation during each step of the loop by incrementing a running index array. It's now (rather remarkably) >5x faster than MAX(DIMENSION=3) for me with large 3D arrays. And of course it gives all the intermediate cumulative max values.
You can easily show how inefficient MAX is on the final dimension with a simple example:
IDL> a=byte(randomu(sd,300,400,3000)*256)
IDL> t=systime(1) & b=max(a,DIMENSION=3) & print,systime(1)-t
IDL> t=systime(1) & b2=a[*,*,0] & for i=1,3000-1 do b2>=a[*,*,i] & print,systime(1)-t
IDL> print,array_equal(b,b2)
I can only presume the built-in MAX has some design limitations for final dimension looping. I presume this all works the same way for MIN, BTW.
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
