Routing problem on new platform import

[marcopoles]

Hi,
I'm creating a new platform to manage tsmc40nm, but I have problem during routing phase: the flow stops due to "heap underflow during 3D maze routing" for the first net to route. The placing phase finish correctly.
I run the flow on gcd and on a very simple verilog with 5 logic ports and on both cases the routing doesn't work.
I've already checked that std cell library pins are aligned with tracks that I defined.

Do you have any suggestion on how to debug the problem? How is it possible to print the path coordinates during the routing operation to check if it starts correctly?

Thanks

[marcopoles]

I found the problem: routing/fastrouting don't manage adjacent metals with the same routing orientation

[eder-matheus]

@eder-matheus we should have a check for layer direction alternation

In this check, should we error out if two adjacent layers have the same direction?

I'm not sure why the direction of a layer below the min routing layer would matter. @eder-matheus any ideas?

If I understand correctly, the underflow occurs when setting M2 as the min routing layer. If using the M1 layer with vertical direction as the min routing layer, the underflow will not occur. @marcopoles Is that right? If so, perhaps the extra resources from M1 helped to avoid the congestion iterations.

Now, if the only change was the M1 direction, but you still using M2 as the min routing layer, it's something new. I'll try to replicate it with our public PDKs and check what is going on.

[maliberty]

@eder-matheus I think it should be an error as we don't handle that case. The check should only apply within the min/max routing layers.

[eder-matheus]

@maliberty I've created a pull request adding a check for adjacent layers direction.

I've also tested with sky130hd setting li1 direction == met1 direction (HORIZONTAL), with min routing layer = met1, and the heap underflow still happens.

This is because we use the first layer's direction to infer the direction of the remaining layers. For example, if li1 (layer 1) is horizontal, the following odd layers are treated as horizontal too, and the even layers are vertical. I can change this to have the direction information directly from the DB instead of using the assumption that adjacent layers have different directions.

[maliberty]

It should come from the db. In any case it should start from the min routing layer rather than the first.

[eder-matheus]

Ok, I'll add this change to the existing PR #4086.