Hi, can we run the openlane flow on multiple cores...
# openlanen
Nguyen Dao
03/31/2021, 1:38 PMHi, can we run the openlane flow on multiple cores? Or is there any option to speed it up?
a
Ahmed Ghazy
03/31/2021, 1:39 PMCurrently you can only set the number of threads to use for detailed routing, which takes the longest, using the
ROUTING_CORESn
Nguyen Dao
03/31/2021, 1:44 PMThanks @Ahmed Ghazy, do you know what is the maximum number can we set (default=4)?
r
Rob Taylor
03/31/2021, 3:52 PM@Nguyen Dao you can ususually use nproc to see your number of processors
Rob Taylor
03/31/2021, 3:53 PMso something like
export ROUTING_CORES=$(shell nproc 2>/dev/null)Rob Taylor
03/31/2021, 3:53 PM@Nguyen Dao or if you can use makefiles, see nproc.mk in openlane top directory
Rob Taylor
03/31/2021, 3:54 PM@Olof Kindgren this is a perfect example usecase for edalize-run-in-cloud
n
Nguyen Dao
03/31/2021, 3:55 PMI only see it runs on 1 cpu when checking the running process (openlane)
Nguyen Dao
03/31/2021, 4:06 PMI set
ROUTING_CORESr
Rob Taylor
03/31/2021, 4:40 PM@Nguyen Dao is it still running?
n
Nguyen Dao
03/31/2021, 4:42 PMI’ve stopped it to change the options, now re-run it but don’t see any improve
r
Rob Taylor
03/31/2021, 4:43 PM@Ahmed Ghazy looks like this could be some kind of pathalogical case or infinite loop. Test case is https://github.com/FPGA-Research-Manchester/FABulous-SKY130
Rob Taylor
03/31/2021, 4:43 PM@mkk ^
o
Olof Kindgren
04/01/2021, 6:26 AM@Rob Taylor Actually had a talk with @Ahmed Ghazy and @mshalan yesterday about Edalize, OpenLane and cloud use-cases
r
Rob Taylor
04/01/2021, 8:43 AMnice :)
n
Nguyen Dao
04/01/2021, 12:54 PMUpdate: the flow works fine with 1x2 and 2x2 CLBs fabrics without timing optimization (OpenPhySyn doesn’t support disable timing loops). For the 4x4 CLBs fabric, it’s stuck at global routing (fastroute)?
r
Rob Taylor
04/01/2021, 1:00 PMDoes the time go up exponentially?
Rob Taylor
04/01/2021, 1:00 PMCould we produce a 2x2 clb gds and then stitch together to make a bigger fabric?
n
Nguyen Dao
04/01/2021, 1:05 PMRunning time are almost same: 12s (1x2 clbs) and 13s (2x2 clbs)
r
Rob Taylor
04/01/2021, 1:06 PMInteresting. Same for 3x3?
Rob Taylor
04/01/2021, 1:07 PMI wonder - are you hitting memory limits and it’s going into swap?
n
Nguyen Dao
04/01/2021, 1:09 PMI haven’t tried 3x3, need to ask Jing to generate codes. I don’t think memory is limited, it used only 1% of memory (64GB) when routing
r
Rob Taylor
04/01/2021, 1:09 PMSame when going the 4x4?
n
Nguyen Dao
04/01/2021, 1:10 PMyes, same memory at 1%
r
Rob Taylor
04/01/2021, 1:11 PMVery odd. Ok let’s get a 3x3 tested, find out where this breaks and then let’s submit a bug
Rob Taylor
04/01/2021, 1:13 PMOr maybe 2x4 if that’s easier
n
Nguyen Dao
04/01/2021, 3:30 PMit’s OK with 2x4, global routing took 16s, looks like the running time doesn’t go exponentially
r
Rob Taylor
04/01/2021, 3:33 PMWeird!
Rob Taylor
04/01/2021, 3:34 PMOk, it’d be good to prepare a test case that shows the working version and the non working version and submit that as a bug against openlane
m
Minh HUST
10/30/2023, 9:36 PMDid you solve the problem? We got the same problem when tried to route small memory.
9 Views