[ERROR GPL-0305] RePlAce diverged during gradient descent calculation

[oharboe]

Describe the bug

untar and run global_place_skip_io_tag_array_64x184_asap7_mocked_2025-07-08_09-04.tar.gz

     2210 |   0.3337 |  0.000000e+00 |   +0.00% |  1.42e+33 |      
     2220 |   0.3337 |  0.000000e+00 |   +0.00% |  2.09e+33 |      
     2230 |   0.3337 |  0.000000e+00 |   +0.00% |  3.07e+33 |      
     2240 |   0.3338 |  0.000000e+00 |   +0.00% |  4.53e+33 |      
     2250 |   0.3337 |  0.000000e+00 |   +0.00% |  6.67e+33 |      
[ERROR GPL-0305] RePlAce diverged during gradient descent calculation, resulting in an invalid step length (Inf or NaN). This is often caused by numerical instability or high placement density. Consider reducing placement density to potentially resolve the issue.
Error: global_place_skip_io.tcl, 12 GPL-0305
openroad> 

Expected Behavior

No error, used to work.

Environment

OpenROAD v2.0-22423-gd59854ee88

To Reproduce

See above

Relevant log output

Screenshots

No response

Additional Context

No response

[oharboe]

Tried a bisection, a090eba is the most recent that supports the .odb format in the test-case and it fails too, so nothing to bisect on with this test-case.

$ git log --pretty=format:'%h' -- OpenROAD | while read commit; do  git show $commit OpenROAD | grep "Subproject commit" | awk '{print $3}'; done | head -n 10
c79c8317e05ecbf45eba5096a74cb421ab97b7ae
a090ebaa03b3171cc7282e49d651c8280d9372b2
...

[gudeh]

Hi @oharboe, was this design working previously or is it a new one? Have you notice the constant 0 HPWL?

looking at it we seem to not have any net. I right clicked and highlighted all nets of all instances and no net showed up:

[oharboe]

Yes: it is mocked SRAM with rtl eviscerated. all I need are blockages and io pins(the .lef file).

[gudeh]

So you would want for gpl to finish the placement even without any nets present? Just avoiding the error, lets say?

[oharboe]

So you would want for gpl to finish the placement even without any nets present? Just avoiding the error, lets say?

Yes.

[gudeh]

I avoided throwing an error with this small modification at my branch: https://github.com/gudeh/OpenROAD/tree/gpl-avoid-error-at-0-hpwl.

Although I am not sure if we want to make this default behavior. What do you think @oharboe and @maliberty ?

I also noticed we have a gpl test named nograd01 which also has 0 HPWL, although it does not throw an error because there so few instances that overflow is 0, and finishes normally, since an overflow below 0.1 is our stop criteria.

[oharboe]

@gudeh Thanks! A warning and avoiding yet another option, i.e. that this just works, would be my preference.

[maliberty]

I think that's fine as it is such an odd corner case. (Strictly you could still resolve density without wire length to consider but I don't think it worth the bother).

[gudeh]

There is more to this issue.

The crash of the design happens due to a division to 0 during gradient calculations. If the HPWL gradient of all instances is 0, we should finish nicely with only the density portion of the gradient, although we don't reach 0.1 overflow. We get stuck with this placement, moving just a little each iteration, and eventually failing:

There is a cluster of 66 instances on top of each other, they seem to have dangling nets connected to only one iTerm, forming this strange force:

At first I thought we had no nets, but we actually do, and I am trying to understand why these instances move tightly together.

Furthermore, PlacerBase and NesterovBase are reporting different HPWL values

[oharboe]

@gudeh Fascinating and pleasing pictures. I presume these are goodies hidden and visible only to those that know what they mean and know how to enable them :-) It is a shame that users can't enjoy them. Probably without knowing exactly what they mean or what to do about the information as it is debug info/visualizations...

[gudeh]

Hi @oharboe, I just used gpl debug mode. I added global_placement_debug to the global_place_skip_io.tcl:

And I ran with GUI activated.

I recently added the option -generate_images to global_placement_debug. There are some info in gpl REAMDE also.

[oharboe]

Nice! It is not the sort of thing that users casually see though...