LabVIEW
Replace Array Subset in FGV slow
Solved!
As a very general comment, please don't attach code as snippets when including called subVIs, because a lot of functionality is stripped out (e.g. terminal assignments, execution settings, etc. etc.).
Please take your toplevel VI and do a "save for previous" (2020 or below) and you'll get a new folder containing all relevant VIs (incl. dependencies) and as a bonus, they'll be in a version many more of us can open..
Zip it up and attach. Thanks! (... and if you do, I'll even have a look at it!)
@Quiztus2 wrote:
I experience a huge performance boost with this version of the fgv.
I am 90% sure this comes from switching from 3D to 1D.
I don't know DVR, but sound interesting.
One possible issue, not sure how smart the compiler is. Run your wires through for all the cases, do not use "Use Default if unwired". Sometimes the compiler is not sure if the buffer will change because now you have an unwired case. Because of that, it may make a copy somewhere which would slow down your FGV.
The replacement using DVR needs < 1 ms in contrast to 40 ms value based.
The reading and creation of the array values again from DVR is slower than value based.
I think, part of your benchmark also included penalties for displaying arrays on the FP.
Just as suggestion - do it like this, this should improve execution time a little bit:
Parallel recreation of native array from DVR is slower than sequential.
For me important is, that my critical usecases:
- Exctraction of single pixel along all images
- Replacement of single image
are significantly faster using DVR. I assume here that the DVR creation from an array doesn't take significant time.
@Quiztus2 wrote:
I assume here that the DVR creation from an array doesn't take significant time.
The DVR method is still the fastest in my quick look at it, although I think the test should be modified like below to include DVR Creation. (Even with the mod, DVR is the fastest.)
The reason the FGV Replace (NON_DVR) is so slow, it appears LabVIEW is making multiple buffer allocations with that method. If you look at the "Profile Buffer Allocations" Tool you will see lots of buffer copies of the data. I believe that is why it is much slower.
Some quick mods on my computer and DVR case is now slower. VI is attached, below are the mods.
The region below in the sequence structure is not needed; it was only to initialize your values to 8888 so the equals works. You can keep if desired.
No need to reshape the array
For my test I included the DVR creation, like in my earlier reply.
