StanfordAHA/garnet

Long RTL generation times

Opened this issue · 9 comments

Runtime for generating the 32x16 array of tiles (with the interconnect only option) has crept back up to ~25 minutes. It was around 5 minutes during the tapeout last year.

That's weird. @leonardt - what was the command you used to profile the python runtime last time?

thanks @leonardt.

@alexcarsello can you run cprofile and paste the output here so we can see what's happening?

@alexcarsello how are we looking on this?

Kuree commented

Just my own datapoint. On kiwi with update-to-date dependencies, 12x12 takes about 7 minutes to generate. Just for comparison, last year 12x12 takes about 1 minutes to generate.

To confirm you're running python garnet.py --width 12 --height 12 --verilog ?

I can c-profile and check that it's not anything silly if it is blocking.

Kuree commented

Yes

@rsetaluri can you try phanrahan/magma#701 and see if it helps?

@leonardt turns out a lot of the time discrepancy is due a gemstone-level issue (not caching properly).

Could be that further magma optimizations could help also. Will try this out, thanks.