Mean Fonction (?)

[ouadji]

from Labview Idea Exchange : @altenbach le ‎08-23-2013 07:01 AM
Please start a regular thread and post your benchmark code.
My guess is that the second code is artificially fast due to constant folding and thus not a valid reference point.
This needs to be discussed in detail elsewhere.

Hello everyone,

Just a small question because I am amazed ...

Why is the Mean_function so slow?

8ms vs 1304ms ???

Thanks to all

[GerdW]

Hi ouadji,

Altenbach already gave you the answer, so what else do you expect here when reposting that question?

See this BD with activated "constant folding" display:

[KathrynB]

If you don't understand what the constant folding means check out the page at http://zone.ni.com/devzone/cda/pub/p/id/347 - so in effect with your benchmark, the dll is being called 10000 times while the code in the loop is only being called once - I think the rates look pretty good to me, in fact, in order to remove the constant folding I called the CONTENTS of the both loops as subVIs - and the difference was then about 60ms on my machine between the two. Without the subVIs I was getting about 120ms on the dll vi (so a sub VI anyway) and 0 on the constant folded maths, put the maths into a subVI (so that it is a fiarer test really as then both are calling another VI) and it jumped up to about 170ms (no more constant folding, going off to somewhere else - a bit like the other one...)

Why is the Mean function so slow - it isn't, in fact in a like for like test (subVI call within a for loop of either the dll using the mean function or a sum array divided by array size - I just highlighted the code and created a subVI) then doing the code by hand is slightly slower on a very simple benchmark test such as the one in the original post

[ouadji]

ok

i didn't know the "constant folding" !

i have disabled the "constant folding" with a shift register, like this :

How to disable "constant folding"? Is it possible to do this? (Maybe in the .ini file?)

[Intaris]

Now if you move the DLL call from within the "Mean.vi" out into the For Loop for testing you'll also lose the sub-VI call overhead fromy our benchmark.

[ouadji]

little or no sub-VI call overhead,

I don't see change, the results are the same. (131/51)

But on the heart of the problem, you are right.

[crossrulz]

Here's my benchmarking (done in 2012).  I am getting a 3x improvement by using the primitives.  I still have some ideas of a few things to try though.

I think the DLL is doing extra checks in there to make sure we don't have an overflow.  That would cause the slow down.

[crossrulz]

I was right.  The DLL code is using an algorithm that avoids overflow to INF.  Here is my final benchmark.  Now, how often would you run into this situation?  I would think it is very rare.  But that is the reason for the time difference.

[Intaris]

Well at least we finally got to a value which we can discuss.

The pitfalls of Benchmarking code are many.

Shane.

[crossrulz]

Ok, one final benchmark in attempt to put this DLL call slow business to sleep.  Here is how I figured NI is doing the average: Sum(X/size(X)).  This is pure speculation, so if somebody can direct me to a better algorithm that avoids the overflow, let me know.