DCplxTrans conversion accuracy

[s910324]

Hi Matthias,

I found while doing DCplxTrans to ICplxTrans conversion the disp loose alot of accuracy, basically the value kind of being rounded to integer first before being convert to dbu unit.

Unlike DTrans to Trans, disp was first divided by dbu then being convert to int, we only loose the accuracy due to the grid size, after DCplxTrans --> ICplxTrans conversion 12.34567 um becomes 12.00 um

Currently I always extract the disp part and convert the value separately to a new ICplxTrans, I was wondering whether there's a suggested way to make this type of conversion?

import pya
dbu = 0.001
t1  = pya.DCplxTrans(1.0, 45.0, False , 12.345678, 20.000000)
t2  = pya.DTrans(12.345678, 20.000000)

print (f"T1 to dbu type: {dcplxtrans.to_itrans(dbu)}")
print (f"T2 to dbu type: {dtrans.to_itype(dbu)}")

>>> T1 to dbu type: r45 *1 12000,20000
>>> T2 to dbu type: r0 12346,20000

[klayoutmatthias]

Hi @s910324,

thanks for reporting this issue. It is actually easy to fix, but I just released 0.28.15, so it did not go there for now.

As a workaround I'd suggest the canonical way, i.e. combining three transformations: one going into micrometer space, then the micrometer-space t1 and a last transformation going back to DBU space:

to_um = pya.CplxTrans(dbu)
from_um = to_um.inverted()
t1_itrans = from_um * t1 * to_um

print (f"(correct) T1 to dbu type: {t1_itrans}")

This gives you:

(correct) T1 to dbu type: r45 *1 12345.678,20000

Side note: You see in that case, that the displacement of the ICplxTrans is even non-integer. Although ICplxTrans transforms integer coordinates to integer ones, internally it uses floating-point representations. This allows implementing "precise" scaling and rotation with rounding only happening on delivery of the transformed coordinates (late rounding). Consequently, the displacement is floating-point too. The effect is that you should get the same results, regardless whether you first transform the coordinates to micrometer and transform it with a micrometer-unit transformation or whether you directly apply the corresponding ICplxTrans.

I also noted the "to_xxx" names are messed up for ICplxTrans and VCplxTrans (e.g. ICplxTrans#to_itrans actually should be to_dtrans). I guess I will clean this up too.

[klayoutmatthias]

0.29 is released now :)