Add CHANZ type

[soheilshahrouz]

Adds CHANZ RR node type for inter-layer connections. Previously, these connection were modeled with CHANX nodes whose ptc number started from the maximum channel width.

[soheilshahrouz]

The RR graph is mostly the same, except some nodes are CHANZ instead of CHANX.

The router lookahead is different from the one in the master branch. Most entries have higher delays. I compared cases where the source and sink are on the same layer, and the lookahead from this branch is more similar to the one from the 2D architecture in the master branch. I think using CHANX for inter-layer nodes may have confused the router lookahead.

[soheilshahrouz]

The problem with the router lookahead is that during RR indexed data initialization, CHANZ nodes are assigned the same RRIndexedDataId as the last CHANX segment type. Many of these CHANZ nodes (which were marked as CHANX) are disconnected from the rest of the graph (fan-in = 0), so their delay is set to zero. This lowers the average delay calculated for that CAHNX wire type.

In this branch, CHANZ nodes are excluded from the average delay calculation for RR indexed data, which makes the CHANX delay estimate more accurate. However, since CHANZ nodes still use the same RR indexed data as CHANX, their delay values are incorrect. As a result, the router lookahead gives inaccurate cost estimates when crossing layers.

[vaughnbetz]

Should that issue (CHANZ rr_indexed_data is wrong) be fixed in this PR too?

[soheilshahrouz]

Should that issue (CHANZ rr_indexed_data is wrong) be fixed in this PR too?

I think a clean fix requires fixing bugs and issues in 3d switch blocks. If the long-term plan is to replace 3d SBs with scatter-gather patterns, I think it's better to postpone fixing the rr_indexed_data until then.

[vaughnbetz]

CI: fail (sigabrt I think) on 3d_sb on reg strong. Should run it through the sanitizers.
QoR: should run one 3D SB arch (can ask Amin) and measure QoR.

[vaughnbetz]

Some changes suggested, plus need to fix the seg fault, plus should run a QoR run on at least one 3D SB arch.

[vaughnbetz]

Could widen the CHANX and CHANY limits too

[soheilshahrouz]

3d_sb_titan_other_auto_bb

	place_time	total_swaps	placed_wirelength_est	placed_CPD_est	route_time	CPD	heap_pops	WL
ratio	0.9863	1.0129	1.0136	1.1523	0.3736	0.9958	0.2993	0.9733

Placement CPD increased because T_linear now gives a more accurate delay for L16 wires in CHANX. Previously, delays for traveling along the x-axis were heavily underestimated.

Route time and heap pops dropped by 63% and 70%. I believe the old router lookahead misled the placement engine by making long CHANX wires seem much faster than they actually were. With a more accurate lookahead, placement timing cost is more aligned with routing timing cost.

[vaughnbetz]

Looks good, thanks.

[vaughnbetz]

Only failure; expected QoR failure on a 3D placement estimated delay. Update and merge!

[soheilshahrouz]

I ran mcnc benchmarks to make sure that adding CHANZ does not impact 2D architectures. I get the same QoR on both branches.