Variants

[oharboe]

Stage: cts

Variant	base	1	2	3	4
crpr	15.07	17.05	19.19	19.19	19.15
skew	519.39	538.16	626.28	626.28	645.81
slack	-4721.79248	-4996.959473	-3951.04834	-3951.04834	-3866.626953
tns	-433454784.0	-458388384.0	-112753840.0	-112753840.0	-107366416.0
MACRO_PLACEMENT_TCL			$(location write_macro_placement)	$(location write_macro_placement)	$(location write_macro_placement)
SETUP_SLACK_MARGIN		0			0
SKIP_CTS_REPAIR_TIMING					0
SKIP_REPORT_METRICS				0
SYNTH_HIERARCHICAL			0	0	0
2_1_floorplan.log	1849	5745	682	1160	3003
2_2_floorplan_io.log	51	39	21	22	21
2_3_floorplan_tdms.log	1	0	0	0	0
2_4_floorplan_macro.log	1728	1365	24	24	23
2_5_floorplan_tapcell.log	43	36	20	20	19
2_6_floorplan_pdn.log	901	791	895	790	897
3_1_place_gp_skip_io.log	1994	1659	2092	1879	1892
3_2_place_iop.log	54	51	32	44	30
3_3_place_gp.log	4720	4523	6473	6643	6111
3_4_place_resized.log	675	652	330	1200	322
3_5_place_dp.log	1682	1527	840	2166	1218
4_1_cts.log	603	584	628	3395	626

Base configuration variables

Variable	Value
CORE_AREA	2 2 1998 1998
DIE_AREA	0 0 2000 2000
FILL_CELLS
GPL_ROUTABILITY_DRIVEN	1
GPL_TIMING_DRIVEN	0
HOLD_SLACK_MARGIN	-200
IO_CONSTRAINTS	$(location :io-boomtile)
MACRO_PLACE_HALO	19 19
MAX_ROUTING_LAYER	M7
MIN_ROUTING_LAYER	M2
PDN_TCL	$(PLATFORM_DIR)/openRoad/pdn/BLOCKS_grid_strategy.tcl
PLACE_DENSITY	0.24
PLACE_PINS_ARGS	-annealing
ROUTING_LAYER_ADJUSTMENT	0.45
RTLMP_FLOW	1
SDC_FILE	$(location :constraints-boomtile)
SETUP_SLACK_MARGIN	-1300
SKIP_CTS_REPAIR_TIMING	1
SKIP_INCREMENTAL_REPAIR	1
SKIP_LAST_GASP	1
SKIP_REPORT_METRICS	1
SYNTH_HIERARCHICAL	1
TAPCELL_TCL
TNS_END_PERCENT	0

[oharboe]

@maliberty @jeffng-or global route is futile for now, it has been running overnight. This makes me wonder: is there an option to have global route give up early?

That would give me a bazel artifact with an .odb file and a .rpt file to examine the problem and it would avoid wasting build resources.

default: -congestion_iterations 30 -congestion_report_iter_step 5 -verbose

Try GLOBAL_ROUTE_ARGS=-congestion_iterations 5 -congestion_report_iter_step 5 -verbose in a seperate PR

[maliberty]

You've already answered your question - use -congestion_iterations with a lower value

[oharboe]

This makes me wonder: is there an option to have global route give up early?

Yes, it was a rethorical question giving myself the answer. Just sharing. I should have worded it more clearly.

[oharboe]

@jeffng-or @maliberty Tried some variants to test out the build system and satisfy some curiosity. No intention to merge.

Some observations:

• WNS is pretty good by the end of floorplan, but something is happening by the time we've finished CTS where WNS goes from -1000 to -4000ps. Not surprising. Merits further study.
• flattened synthesis with macro placement from synth_hierarhical=1 run and CTS repair is enabled in variant 4 and WNS after CTS is best in this case -3800ps
• from variant base to 1, SETUP_SLACK_MARGIN is modified from -1300 to 0, but WNS after CTS is very slightly worse.
• between variant 2 and 3, SKIP_REPORT_METRICS=0 is enabled. This increases running times, obviously. Since SKIP_REPORT_METRICS is enabled whether or not it is useful for substages, I think the running times for reports could be best mitigated by having an option to run it only at the end of a stage, rather than for each substage.
• between base and 1 variant, SETUP_SLACK_MARGIN is changed from -1300 to 0, but there is no change, essentially, in WNS after CTS.
• cts running times between variant 2 and 3 goes from 600 to 3000 seconds. This is all reporting metrics.

[oharboe]

@maliberty @jeffng-or Interesting: looks like macro placement fails with SETUP_SLACK_MARGIN=0 instead of -1300 when I no longer dissolve the register files. I imagine that SETUP_SLACK_MARGIN=0 can add a lot of instances and make macro placement fail. Will look into creating a standalone reproduction case.

[oharboe]

@maliberty The WNS is getting worse in repair timing in CTS before it ends up roughly where it started. Is that interesting as a test-case or is it to be expected?

[oharboe]

@jeffng-or @maliberty Something bad is happening during blobal placement? Could it be the routing that is wrapped several times around a macro?

$ grep ws $(ls bazel-bin/logs/asap7/BoomTile/3/*.json | sort -h)
bazel-bin/logs/asap7/BoomTile/3/2_1_floorplan.json:	"floorplan__timing__setup__ws": -1350.4,
bazel-bin/logs/asap7/BoomTile/3/3_3_place_gp.json:	"globalplace__timing__setup__ws": -14623.4,
bazel-bin/logs/asap7/BoomTile/3/3_4_place_resized.json:	"placeopt__timing__setup__ws": -3492.46,
bazel-bin/logs/asap7/BoomTile/3/3_5_place_dp.json:	"detailedplace__timing__setup__ws": -3477.23,
bazel-bin/logs/asap7/BoomTile/3/4_1_cts.json:	"cts__timing__setup__ws": -3517.24,

[oharboe]

@jeffng-or @maliberty This PR was an experiment to learn something and file some github issues with feature requests and bugreports.