Jump to content


Popular Content

Showing content with the highest reputation on 01/07/2019 in all areas

  1. 1 point
    Don't you use a modified OpenGPackage Manager for zlib rather than VIPM? I'm not sure if it is the same issue but when VIPM compiles the *.* it "fixes" the file path to whatever bitness you compiled it in if it successfully finds the library. That means if you created the package with 64 bit LabVIEW and the file name is "something_*.*", then the path gets "fixed" to "something_64.dll". The result is that something_32.dll isn't searched for when installed and recompiled on a 32 bit platform and the Vis are broken.
  2. 1 point
    My general advice for refactoring code is to do it step-by-step. Don't try to fix everything at once, as it will likely break your code (as it did in your case). Note: I wasn't able to load your VI snippet. I'm not sure why though (either I'm to stupid to do it or Lava sanitizes images) My suggestions below are based on your original VI Step 1 - Improve readability Arrays and for loops The outer for loop iterates over multiple arrays at the same time. This means that the loop will end after processing each element of the smallest array. I suggest you use one array for iteration and use the Array Index function to access elements from the other arrays. Here is an example: You should play around with it to see the different outcomes. Note that indexing an element from an array of smaller size returns a default value (i.e. zero for numbers, an empty string for Strings and false for Boolean). Shift registers The outer for loop uses a shift register for the SQLite instance as well as the error cluster. The inner for loop, however, doesn't. This can be problematic for two reasons: 1) The SQLite instance in the inner for loop gets reset after each iteration. 2) An error in the inner for loop is reset after each iteration and only the last error is returned. The solution depends on what you want to do. If you want to abort as soon as possible on any error, I suggest not using a shift register but making the for loop conditional instead (exit on error). Otherwise use shift registers on both loops. Step 2 - Optimize your code Use transaction outside of for loops Transactions are very useful for improving performance if used correctly. In your second example, you already added the BEGIN and COMMIT Transaction using "MySavepoint". This is generally a good idea. Disable transactions on Execute SQL The Execute SQL function has an optional input on the top which you can set to false in order to disable transactions. This will already improve performance considerably. Answering questions SQL Prepare and SQL Finalize should be placed outside the for loop and you need to inject values for your parameters at some point (unless your SQL Statement is static, which it isn't according to your info above). The "INSERT many rows Template code.vi" provided by the library (see Code Templates palette) visualizes that very well: I highly suggest not using "SELECT *" but explicitly selecting particular columns instead. That way your SQL Statement becomes predictable. Here is an example that may or may not work (I can't test it for obvious reasons): I suggest making a VI for each statement (without transactions of course). That way you can test things in isolation and be certain that it works. I'm not sure what you mean by that. Can you explain what you mean in more detail?
  3. 1 point
    I"m not sure it's conflation only. Many seem to be focused specifically on the fact that LabVIEW GUIs don't look like the latest hyped Office version, which of course will be again different over 2 years when yet another modern style design guide claims that everything needs to be high contrast again, or maybe alpha shaded with psychadelic animations (you need to find reason to sell high end GPUs to spreadsheet jugglers). Your second point is indeed one thing I feel LabVIEW could have made more advances. Dynamic control creation while indeed complicated to get into the dataflow paradigma of LabVIEW would be possible with the VI Server reference model, although it would not be classical dataflow programming anymore for such GUIs at least for the part that you use such dynamic instantiation in. XControls was a badly executed project for an in principle good idea, Splitters are a God send for more dynamic UIs together with Subpanels, but to try to edit splitters once they are placed on a frontpanel really can be an exercise in self control. In doing so I feel the same frustration as when I'm forced to edit some Visio drawings. The editor almost seems to know what you want to do, because it always does the opposite of that, no matter what.

  • Create New...

Important Information

By using this site, you agree to our Terms of Use.