scany | scanning data from a database into Go structs | SQL Database library

you're hitting the limitations of scripting languages here. Using sliceN is already pretty much the optimum you can get to, unfortunately. Everything else in speed optimization requires parallelized, compiled code. (i.e. you could code C++ code and use the SDK to compile your own plugin.)

However, there is a bit of room for improvement over your example.

First of all, your example above doesn't run :c) But that is quickly fixed.

Try to avoid number type casting. DM script only knows number but internally there is a difference between the proper number types (integer, floating point, signed/unsigned, byte-size). The script languages uses real-4-byte as the default unless told differently explicitly. And some methods will return real-4-byte by default. For this reason, the processing will be fastest, if both data and mask use real-4-byte data as well.

In my testing, the time-difference between running with uint16 data plus uint8 mask and *real4 data plus real4 mask) was significant! Nearly 30% time difference.

Don't copy you sliced image! Use := not = for your Dif2D.

The SliceN command returns an expression directly addressing the required memory. You can use it directly in any other expression (like I do below) or you can assign an image variable to it using := to give it a name.

The speed increase is not huge, but it's one copy-operation less per loop iteration.

You additional knowledge: Now for arbitrary masks there is not much you can do, but most often masks are zero-valued over large stretches and it is possible to define a smaller ROI containing all non-zero points. If this is the case, you can limit your math operations to that region.

i.e. instead of multiplying the whole DP with the same sized mask, just use a smaller mask and use the according sub-section of the DP.

This can actually make a big difference, but it will depend on your mask.

Of course you need to "find" this ROI first. In my script below I'm having a helper method to do that, utilizing the comparatively fast max() command and image rotation as trick for speed-up.

...would be to get rid of the double-for loop and replace it with image-expressions. Unfortunately, DigitalMicrograph does currently (GMS 3.3) not support this for 4D or 5D data.

The script below executed on a [53 x 52 x 512 x 512] STEM DI (of real-4 byte data) gave me the following timings: