Inverse of the average number of tracks per channel

[soheilshahrouz]

Before this PR, calls to std::min() and std::max() clipped bounding box coordinates between 1 and grid width/height - 2. This PR removes all these calls and initalizes chanx_place_cost_fac_ and chany_place_cost_fac_ in a way that is compatible with the new index range: [-1, grid width/height - 1].

[soheilshahrouz]

QoR is compared with #2676

This is measured before inlining operator() of ChanPlaceCostFacContainer class.

Titan Quick QoR	Pack Time	Place Time	Place WL	Place CPD	Place Mem	N Swaps
net_cost_handler	653.0400675	1268.961831	2496205.323	17.12522422	5491.752447	18117190.22
chan_fac	655.37998	1270.253766	2496205.323	17.12522422	5491.853519	18117190.22
ratio	1.0036	1.001	1	1	1.000018	1

[soheilshahrouz]

/vtr_reg_nightly_test3/vtr_reg_qor_chain_large
QoR is compared with #2676
This is measured after inlining operator() of ChanPlaceCostFacContainer class.

VTR Large	Pack Time	Place Time	Place WL	Place CPD	Place Mem	N Swaps
vtr_net	6.43E+01	69.57212354	231158.1631	20.42328639	685.6120645	2080613.564
vtr_chanxy	6.49E+01	69.04388726	234105.7188	19.79841244	685.3785724	2090490.813
	1.01E+00	0.9924	1.01275	0.9694	0.9997	1.0047

[soheilshahrouz]

QoR is compared with #2676

This is measured after inlining operator() of ChanPlaceCostFacContainer class.

Test Case	Pack Time	Place Time	Place WL	Place CPD	Place Memory	Number of Swaps
titan_net_inline	650.514155	1269.200989	2496205.323	17.12522422	5491.663657	18117190.22
titan_chanxy_inline	653.6507028	1260.628876	2496205.323	17.12522422	5491.771652	18117190.22
ratio	1.004821644	0.993246056	1	1	1.000019665	1

[vaughnbetz]

The code all looks good -- I have no comments to address :).

[vaughnbetz]

Whoops, I lied. I think the [-1, ...] indexed Matrix you want can just use the VtrNdOffsetMatrix, and doesn't need a new class.

[vaughnbetz]

Ah, I see. Maybe then the new class you've created should be pulled out into a new utility class that allows -ve offsets on top of a VtrNdMatrix?

[vaughnbetz]

Thanks! The changes to generalize and use nd_offset_matrix look good.

[vaughnbetz]

Looks like there are some memory faults / assertion failures. From the valgrind run:

Command being timed: "valgrind --leak-check=full --suppressions=/home/runner/work/vtr-verilog-to-routing/vtr-verilog-to-routing/vpr/valgrind.supp --error-exitcode=1 --errors-for-leak-kinds=none --track-origins=yes --log-file=valgrind.log --error-limit=no /home/runner/work/vtr-verilog-to-routing/vtr-verilog-to-routing/vtr_flow/scripts/blackbox_latches.pl --input 1_stereovision3.abc.blif --output stereovision3.abc.blif --vanilla"

2024-09-26T17:27:08.5689907Z User time (seconds): 0.48
2024-09-26T17:27:08.5690506Z System time (seconds): 0.03
2024-09-26T17:27:08.5691076Z Percent of CPU this job got: 99%
2024-09-26T17:27:08.5691747Z Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.52
2024-09-26T17:27:08.5692402Z Average shared text size (kbytes): 0
2024-09-26T17:27:08.5692996Z Average unshared data size (kbytes): 0
2024-09-26T17:27:08.5693605Z Average stack size (kbytes): 0
2024-09-26T17:27:08.5694154Z Average total size (kbytes): 0
2024-09-26T17:27:08.5694725Z Maximum resident set size (kbytes): 161656
2024-09-26T17:27:08.5695506Z Average resident set size (kbytes): 0
2024-09-26T17:27:08.5696097Z Major (requiring I/O) page faults: 0
2024-09-26T17:27:08.5696702Z Minor (reclaiming a frame) page faults: 7996
2024-09-26T17:27:08.5697324Z Voluntary context switches: 1
2024-09-26T17:27:08.5697839Z Involuntary context switches: 4
2024-09-26T17:27:08.5698355Z Swaps: 0
2024-09-26T17:27:08.5698725Z File system inputs: 0
2024-09-26T17:27:08.5699186Z File system outputs: 80
2024-09-26T17:27:08.5699745Z Socket messages sent: 0
2024-09-26T17:27:08.5700229Z Socket messages received: 0
2024-09-26T17:27:08.5700720Z Signals delivered: 0
2024-09-26T17:27:08.5701175Z Page size (bytes): 4096
2024-09-26T17:27:08.5701620Z Exit status: 0
2024-09-26T17:27:17.4065810Z /home/runner/work/vtr-verilog-to-routing/vtr-verilog-to-routing/libs/libvtrutil/src/vtr_ndoffsetmatrix.h:389 operator[]: Assertion 'this->dim_size(0) > 0' failed (Can not index into size zero dimension).
2024-09-26T17:27:17.6700617Z Command terminated by signal 6
2024-09-26T17:27:17.6706561Z Command being timed: "valgrind --leak-check=full --suppressions=/home/runner/work/vtr-verilog-to-routing/vtr-verilog-to-routing/vpr/valgrind.supp --error-exitcode=1 --errors-for-leak-kinds=none --track-origins=yes --log-file=valgrind.log --error-limit=no /home/runner/work/vtr-verilog-to-routing/vtr-verilog-to-routing/vpr/vpr k6_frac_N10_frac_chain_mem32K_40nm.xml ch_intrinsics --circuit_file ch_intrinsics.pre-vpr.blif --min_route_chan_width_hint 44"

[vaughnbetz]

Update: seg fault fixed, @soheilshahrouz gathering updated QoR to be safe.

[vaughnbetz]

@soheilshahrouz : if you post the QoR data (and it looks good) we can merge this.

[soheilshahrouz]

titan_quick_qor	pack-time	place time	place WL	place cpd	place_mem	n_swaps
master	501.2210495	1183.176432	2496483.069	17.05656962	5501.755819	18137699.29
PR	510.2875014	1206.429456	2496483.069	17.05656962	5501.687095	18137699.29
ratio	1.018	1.019	1	1	0.9999875	1

[vaughnbetz]

LGTM!

[AlexandreSinger]

@vaughnbetz Sorry for all the messages; but I think this PR was the one that is causing the CI to fail now. The CI run for this PR was a week old and since then there have been golden result updates and other minor changes which seems to have been revealed by this PR.

@fkosar-ql Sorry for wrongly accusing your PR!

I am currently regenerating the golden results for the Strong tests since that is easy to fix. The issue with the basic test is probably something I cannot easily fix. Issue #2754 is tracking the issue.

Edit:
see PR #2768