DRC: drc(width) much slower than width check #1195
Assignees
Labels
Comments
klayoutmatthias
added a commit
to klayoutmatthias/globalfoundries-pdk-libs-gf180mcu_fd_pr
that referenced
this issue
Nov 29, 2022
1. No input merging in deep mode Deep mode benefits from a hierarchical distribution of large polygons across the hierarchy (e.g metal1). So merging does not happen in deep mode. (BTW: the same question is not easy to answer for tiled mode. I'd rather opt not to merge at all as this happens internally anyway). 2. drc(width < ...) replaced by width See issue KLayout/klayout#1195 In the future maybe drc will have same performance than width. 3. Precomputation of some redundant expressions e.g. contact_logic = contact.outside(sramcode) use a couple of times This patch by far does not capture all the redundant expressions. There is some room for improvement. I addressed some dominant ones. 4. chosing enclosed instead of enclosing (CO.6) a.enclosing(b) is equivalent to b.enclosed(a), but may be more efficient. Here: contact.enclosed(metal1) is more efficient than metal1.enclosing(contact). Reason: the first argument of all operations usually determines the hierarchical scope and lookup region. contact is more localized and more sparse than metal 1, so traversing contact hierarchically is easier (basically only standard cells to visit) and collecting the neighborhood only means looking up a small region of the layout for interacting shapes. That makes this operation more efficient with contact as the first argument. BTW: same is true for the symmetric operations such and "and" or "separation". 5. Avoid long "or" chains e.g. this: plfuse.not_outside(comp.or(nplus).or(esd).or(sab).or(resistor).or(metal1).or(metal2)) is more efficiently written as: plfuse.not(plfuse.outside(comp).outside(nplus).outside(esd).outside(sab).outside(resistor).outside(metal1).outside(metal2)) Reason: The "outside" operations will gradually reduce the complexity of the layer. For no plfuse present, the outside operation is essentially a null operation. So the argument to "not" will compute very efficiently while the above "or" chain will first compute a very complex region just to discover we need little or nothing of it. BTW: "or" is a real polygon merge and usually is more expensive than required. An alternative is "+" which just joins the polygon collections. 6. Rewriting MDN.3a, 3b, 4a and 4b. The main goal is to avoid long-distance width interactions. The basic idea is to consider the real gate area of the LDMOS device which is "(poly & comp) - mvsd". This is the basic region to consider as it defines the channel. A basic width check on these rectangles is easy and allows to quickly filter out devices outside the range of interest. Channel width and length can be differentiated by looking at edges inside comp or poly2. Note that "inside_part" and "outside_part" deliver edges which are propagating inside the second argument (a region) or entirely outside the region. In contrast to "and" and "not" which will treat edges on the region boundary as "inside". 7. Rewriting MDP.1, 1a and 2. See 6. for the motivation and technique. 8. Adapting "regression.py" This was required as "deep" is used as keyword in the "merged" decision (see 1.). 7. Rewriting MDP.1, 1a and 2. See 6. for the motivation and technique. 8. Adapting "regression.py" This was required as "deep" is used as keyword in the "merged" decision (see 1.). 7. Rewriting MDP.1, 1a and 2. See 6. for the motivation and technique. 8. Adapting "regression.py" This was required as "deep" is used as keyword in the "merged" decision (see 1.). 7. Rewriting MDP.1, 1a and 2. See 6. for the motivation and technique. 8. Adapting "regression.py" This was required as "deep" is used as keyword in the "merged" decision (see 1.). 7. Rewriting MDP.1, 1a and 2. See 6. for the motivation and technique. 8. Adapting "regression.py" This was required as "deep" is used as keyword in the "merged" decision (see 1.). 7. Rewriting MDP.1, 1a and 2. See 6. for the motivation and technique. 8. Adapting "regression.py" This was required as "deep" is used as keyword in the "merged" decision (see 1.). 7. Rewriting MDP.1, 1a and 2. See 6. for the motivation and technique. 8. Adapting "regression.py" This was required as "deep" is used as keyword in the "merged" decision (see 1.).
|
There is a connection with this issue: #1215 See comments below there and a test case. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Test case from: #1189
The following code:
is much slower than
(~60s vs. 9s)
The text was updated successfully, but these errors were encountered: