Inefficient access of structure array within a structure

[rsiddans]

GDL is very very slow (compared to IDL) at accessing individual elements of a large structure array held within another structure.
I.e. this
for i=0,n-1 do m1=x.m(i)
Is drastically slower than
m=x.m & for i=0,n-1 do m1=m(i)

It is as if GDL copies the entire (sub)structure array before accessing the individual structure, rather than directly access the element itself.

Code to test this...

pro test_strarr_access
; make a big array of structures
        m1={a:10}
        n=50000l
        m=replicate(m1,n)
; Time accessing each structure in turn
        t0=systime(/jul)
        for i=0,n-1 do m1=m(i)
        t1=systime(/jul)
        print,'Time to access m: '+string((t1-t0)*24*60*60)+'s'
; Time accessing each structure in turn when the array is inside another structure
        x={m:m}
        t0=systime(/jul)
        for i=0,n-1 do m1=x.m(i)
        t1=systime(/jul)
        print,'Time to access m in structure: '+string((t1-t0)*24*60*60)+'s'
end

Output with IDL and GDL

GDL> test_strarr_access
Time to access m: 0.0053107738s
Time to access m in structure: 22.644134s

IDL> test_strarr_access
% Compiled module: TEST_STRARR_ACCESS.
Time to access m: 0.0037014484s
Time to access m in structure: 0.0088512897s

Note the behaviour of GDL is much better for a simple integer array - e.g. defining "m1=10" instead of "m1={a:10}". The issue seems specific to structure arrays.

[alaingdl]

Thanks you for this clear and useful report. I test it and got very similar figure 😢

[GillesDuvert]

see also #1540
This is an interesting problem. I've looked at the GDL "object" named "Structure" in the past and concluded it would be difficult to improve speed without rewriting completely the object. Will look into that.