Improve global router for asap7/ethmac_lvt where it fails with a single DRC error

[oharboe]

Description

In The-OpenROAD-Project/OpenROAD-flow-scripts#2694 asap7/ethmac_lvt fails global route with a single DRC error after 100 iterations.

To reproduce untar and run https://drive.google.com/file/d/1XCfVGa0DHAz5MO_pJwR_VE8qOhZ1oqoO/view?usp=sharing

No luck in whittling down the problem with deltaDebug.

Suggested Solution

Improve global router.

Additional Context

No response

[maliberty]

The IO pin placement is totally different than master which has few pins on the e/w edges.

[oharboe]

The IO pin placement is totally different than master which has few pins on the e/w edges.

This is unsurprising to me: a completely different optimal solution can be found with tiny differences in initial conditions.

So what does this mean practically?

• Pin ppl chooses a pin placements that is never going to work in global routing and it should have done better, so fix ppl?
• There's nothing wrong here with ppl, nor grt, just lock down pin placement in ORFS test cases in git to avoid random changes breaking master?
• Wait for initial conditions to change again slightly, then merge synth: nudge .rtlil towards cleaner and more canonical OpenROAD-flow-scripts#2694 ?

I ran out of imagination w.r.t. suggesting alternatives :-)

[maliberty]

I think it is worth a bit of digging to see if this is just the flap of the butterfly's wings or something more interesting.

[oharboe]

I think it is worth a bit of digging to see if this is just the flap of the butterfly's wings or something more interesting.

I suppose the test case is somewhat interesting independent of The-OpenROAD-Project/OpenROAD-flow-scripts#2694

[eder-matheus]

@oharboe @maliberty I believe this is a case where we apply an unnecessarily large reducement in the routing resources. We're reducing it in 50% for every asap7 design, with few exceptions. I think this could be the case of studying different reduction values for this platform and choose a better one.

The attached test case finishes if I reduce the adjustment from 50% to 40%.

[oharboe]

@eder-matheus Would it be possible to make the error message more actionable and specific?

"grt failed, try increasing X to Y"?

Even if it takes a while to compute "Y" here, I think it would be worth the extra compute, because all the user can do is to randomly change variables and hope for the best...

[eder-matheus]

@eder-matheus Would it be possible to make the error message more actionable and specific?

"grt failed, try increasing X to Y"?

Even if it takes a while to compute "Y" here, I think it would be worth the extra compute, because all the user can do is to randomly change variables and hope for the best...

I created this issue to track this request: #6638.
And I believe this other issue is also related to this topic: #6404.

I will work on it in the next days, hopefully we can have at least a rough measure for new adjustment suggestion.

[maliberty]

The best we can do is try to guess at the better value. We can't really guarantee anything as you would have to re-run the flow run global placement forward to see if it works.

[oharboe]

my guess is completely uniformed...

All I can do is to scan values and look at the results to try to reason what constitutes a "good" value.