Reducing size of mock abstracts

[oharboe]

Mock abstracts is where cts and route is skipped such that an abstract can be used to more quickly investigate higher level floorplanning, placement and routing.

Today this includes skipping cts_fill, which saves size and speeds things up.

I was wondering if there are any further easy optimizations possible?

After make skip_cts skip_route generate_abstract, I have the files below.

skip_cts and skip_route are targets that does what it says on the tin: skip those stages, mocking the next stage by simply copying the files:

OpenROAD-flow-scripts/flow/Makefile

Line 681 in 51cbbec

.PHONY: skip_cts

There doesn't appear to be any stages that can be safely skipped without creating problems later on.

The PDN stage seems important for a mock abstract as the PDN can interfere at a higher level... global placement is necessary to get the pin layout optimized.

Perhaps 3_4_place_resized.odb could be skipped to save some space?

Could 3_5_place_dp.odb be skipped to save time?

When I say size, of course symbolic links could be used, but really care about execution speed of e.g. deltaDebug.py or OpenROAD when parsing the files.

$ du -hs results/asap7/FooBar/base/*.odb
93M   results/asap7/FooBar/base/2_1_floorplan.odb
94M   results/asap7/FooBar/base/2_2_floorplan_io.odb
94M   results/asap7/FooBar/base/2_3_floorplan_tdms.odb
94M   results/asap7/FooBar/base/2_4_floorplan_macro.odb
95M   results/asap7/FooBar/base/2_5_floorplan_tapcell.odb
107M  results/asap7/FooBar/base/2_6_floorplan_pdn.odb
107M  results/asap7/FooBar/base/2_floorplan.odb
107M  results/asap7/FooBar/base/3_1_place_gp_skip_io.odb
107M  results/asap7/FooBar/base/3_2_place_iop.odb
178M  results/asap7/FooBar/base/3_3_place_gp.odb
184M  results/asap7/FooBar/base/3_4_place_resized.odb
184M  results/asap7/FooBar/base/3_5_place_dp.odb
184M  results/asap7/FooBar/base/3_place.odb
184M  results/asap7/FooBar/base/4_1_cts.odb
184M  results/asap7/FooBar/base/4_2_cts_fillcell.odb
184M  results/asap7/FooBar/base/4_cts.odb
184M  results/asap7/FooBar/base/5_1_grt.odb
184M  results/asap7/FooBar/base/5_2_route.odb
184M  results/asap7/FooBar/base/5_route.odb
184M  results/asap7/FooBar/base/6_1_fill.odb
184M  results/asap7/FooBar/base/6_final.odb

[maliberty]

Since we use 'write_abstract_lef -bloat_occupied_layers' today only a single shape is required to block an entire layer. We could skip PDN and either

• Put a single shape on each layer that PDN would use
• Add an option to write_abstract_lef to give a set of layers to block (nicer)
• Run PDN but skip the vias which probably are the bulk of the data.

If in the future we drop -bloat_occupied_layers then this wouldn't apply.

Dropping resizer would affect the timing model generated. Maybe that doesn't matter given the other simplifications.