Generating SRAM fails at supply_router: Couldn't route from (vdd layer=m3 ...) to to (vdd layer=m3 ...) #268

Activity

1. mguthaus commented on May 14, 2025

   

   mguthaus

   on May 14, 2025

   Collaborator

   More actions

   It's an odd, and very large, word size. I'd consider banking or other optimization.

   Often this type of thing is just caused by a pin access problem, but you would need to identify it in the temp output that is saved. The top of the run should tell you what the GDS file name is to investigate.

   Sorry, but we don't have resources to fix bugs, but would welcome any improvements as a PR>
2. ManzoniGiuseppe commented on Dec 9, 2025

   

   ManzoniGiuseppe

   on Dec 9, 2025

   More actions

   In case anyone wants to try fix it, a similar error also happens with freepdk45, and with the following power-of-two word_size:

   • 128 (it fails with num_words = 512, but succeeds with 256)
   • 256 (it fails with num_words = 64,128, but succeeds with 16)
   • 512 (it fails with num_words = 512)

   For example, this configuration:

   tech_name = "freepdk45"
   word_size = 128
   num_words = 512
   write_size = 8
   num_rw_ports = 0
   num_r_ports = 1
   num_w_ports = 1
   

   Fails with:

   ** Submodules: 1417.4 seconds
   ** Placement: 0.2 seconds
   ERROR: file signal_escape_router.py: line 70: Couldn't route from (addr1[8] layer=m3 ll=v[359.485,-3.61] ur=v[359.62,-3.475]) to (addr1[8] layer=m4 ll=v[359.4825,-15.52] ur=v[359.6225,-15.38]).
   

   I can use smaller SRAMs in parallel, so it's not the end of the world, but it's not as efficient as a single bigger SRAM.
3. mguthaus commented on Oct 13, 2026

   

   mguthaus

   on Oct 13, 2026

   Collaborator

   More actions

   Still open, but worth retesting against dev first: the supply router has changed since you filed this. #314 and #317 both landed on dev and touch exactly this path — #317 replaced recursion in the router's bbox tree, which previously blew Python's recursion limit on large macros, and #314 fixed a graph bug where remove_all_neighbors() skipped every second neighbor, leaving edges into blocked positions. Neither was aimed at a Couldn't route failure specifically, but both change what the router sees.

   If it still fails on dev, the routing code is in compiler/router/, and running with -v -v gives the per-net routing detail.

   If you would like to take it, a PR against dev would be very welcome and we will review and consider it — see CONTRIBUTING.md. Contributions go to the dev branch, and the regression suite is run with cd compiler/tests && make -k -j<N>.