Time to call a dll !!!

[ouadji]

Below, just an example that uses the "mean function".

This function uses a call to a dll.

Look at the difference between the execution times !

This is just a small example to show the extreme slowness of dll calls

The time required to call a dll is a disaster!

                  please, It would be necessary to improve time to call a dll, really !

sorry for my bad english, i do my best

[Intaris]

If someone can post in LV 2012, I'll make tests also and put this topic to rest

PS place the DLL call from the Mean.VI directly in the loop instead of the full VI.

[Mads]

I have to admit my initial test had a flaw as well 

It seems that constant folding is a bit trickier to avoid than expected. If the input array is kept the same then the compiler seems to be so smart that it skips anything but the first run...

To force the same DLL free code to repeat itself each time I did a test where the input array is regenerated on each iteration. This requires subtraction of the time it takes to generate the arrays (varies, but not too much), but seems to remove the constant folding:

On my computer the non-DLL version runs about 2 times faster than the dll-version(s):

Hopefully I've not overlooked another flaw this time  (?) 

[crossrulz]

Please send all other benchmarking to the thread in the LabVIEW board.  It turns out that the Mean code by NI uses an algorithm that is setup to avoid overflow to INF.  That is why it is slower.  It isn't the DLL call itself.

[Intaris]

And we've gone from a difference of 150x to "only" 2.5x

[AristosQueue (NI)]

It appears that the explanation to the slowdown has been found and it has nothing to do with DLL call speed. Is there any reason to keep this idea open longer?

[ouadji]

"Is there any reason to keep this idea open longer?" ... no !

Thank you all.
I didn't know the "Constant Folding", today I learned something!
a difference of 150x to "only" 2.5x ... yes, I agree, my last tests give the same results.
A very interesting thread, really.

[altenbach]

> a difference of 150x to "only" 2.5x.

Actually, the dll is 6x faster when comparing identical algorithms, so it is a difference of 150x to 0.17x. 😄

[Intaris]

Well seeing how we don't actually KNOW what the DLL is doing, that seems to be a rather unqualified statement.....

Besides, I was simply referring to the bad benchmarking mistakes which led to completely wrong initial comparison values.

[altenbach]

Yes, we don't know the dll algorithm, but (as others have pointed out) it seems quite sophisticated when dealing with pathological data. Attempting to duplicate all these features in plain G will most likely slow it down even more, making the dll even more impressive. 😄

Just guessing of course. Maybe there is a mathematical shortcut that I don't know about.... 😮

[JordanG]

Issue was resolved in comments below