# Known Issues ## Cadence Macro Cases Can Crash With OpenROAD-Style Additional Layers ### Summary In Cadence 3D flow, macro-containing designs can crash during `cds-place-upper` when the technology/view stack includes the OpenROAD-style additional routing layers such as `M2_add` / `M3_add`. The current evidence points to an Innovus internal optimizer failure triggered by the combination of: - macro-containing designs - `place_opt_design`-based upper placement - OpenROAD-style additional-layer setup in the Cadence flow This is currently treated as a flow compatibility issue, not a normal design violation. ### Affected Scope - Cadence flow - macro-containing cases such as `swerv_wrapper` - especially `nangate45_3D` mixed-view runs like `cds_ordtech` - stage: - `cds-place-upper` ### Typical Symptoms - repeated internal asserts in Innovus: - `coeTransform.hpp:2625:getTopoLaCount` - final internal crash: - `Innovus terminated by internal (SEGV) error/signal` - trailing worker-thread stacks stuck in `pthread_cond_wait` - this is a post-crash symptom, not the root cause ### Evidence Relevant logs: - [3_place_upper.upper-only.log](logs/nangate45_3D/swerv_wrapper/cds_ordtech/3_place_upper.upper-only.log) - [3_place_upper.upper-only.log](logs/nangate45_3D/swerv_wrapper/cds_ordtech_nosplit/3_place_upper.upper-only.log) Common indicators seen in both split and no-split runs: - additional-layer routing setup is active: - `setDesignMode -topRoutingLayer M3_add` - additional-layer pitch warnings appear: - generated pitch mismatch on `M2_add` - extra tech inconsistency warning appears: - `IMPRM-132 Missing VIARULE GENERATE definition between M19/M20` - later, the optimizer crashes in the same internal topo path: - `getTopoLaCount` - `coeSetupOptimizer` - `rdaOptDesignCL::run` The same signature appears in both: - `cds_ordtech` - `cds_ordtech_nosplit` So this issue is not currently attributed only to split-buffer insertion. ### Current Interpretation The current working interpretation is: - pure `split_net` is not the primary root cause - the main trigger is the Cadence upper placement optimization flow running on a macro-containing design under the OpenROAD-style additional-layer setup - this likely exposes an Innovus internal bug or unsupported flow combination ### Current Workaround Until this is fully fixed, avoid using the OpenROAD-style additional-layer setup for Cadence macro cases in `place-upper`, or reduce the stage to a safer placement mode. Practical mitigations: - avoid the additional-layer tech/view mix for macro cases - avoid `place_opt_design` on the affected macro cases when possible - prefer a simpler `place_design`-style upper placement for debug ### Notes - This issue has only been clearly reproduced on macro-containing Cadence cases so far. - Non-macro cases may still run with the same additional-layer setup. - If this issue is revisited later, compare against: - macro-free Cadence cases - the same macro case without additional layers - the same macro case with reduced placement optimization effort ## Cadence Legacy Route Can Create Large DRC Regressions After `routeDesign` ### Summary In the Cadence legacy route flow, applying: - `apply_tier_policy [cts_owner_tier] -fixlib 1 -allow_net $effective_allow_net` after `routeDesign`, and then running: - `optDesign -postRoute -outDir $REPORTS_DIR -prefix route_legacy` can introduce a large DRC regression. This is currently treated as a flow issue in the legacy route sequence, not as a normal expectation of the route stage. ### Affected Scope - Cadence flow - legacy single-stage route: - `cds-route` - script: - [innovus_3d_route_legacy.tcl](scripts_cadence/innovus_3d_route_legacy.tcl) ### Symptom The route stage completes, but final DRC is much worse than the pure route stage expectation. Observed example: - [final_summary.txt](logs/asap7_3D/swerv_wrapper/cadence_route_legacy_cmp/final_summary.txt) - `DRC Violations = 50466` The corresponding route log shows that the legacy flow explicitly runs owner-tier postRoute optimization after `routeDesign`: - [5_route.log](logs/asap7_3D/swerv_wrapper/cadence_route_legacy_cmp/5_route.log) - `Running optDesign -postRoute ...` ### Current Interpretation The suspected mechanism is: - `routeDesign` finishes on one routed topology - `apply_tier_policy` is then re-applied after route - this changes optimization legality / allowed cell usage / row policy late in the flow - `optDesign -postRoute` then perturbs the routed design under a different policy than the one used during the actual route stage This late policy switch is currently considered unsafe for DRC stability. ### Current Workaround Prefer either: - pure route-only flow for cleaner routing behavior - or staged postRoute flow with explicit receive/owner separation Avoid using the legacy postRoute step as the baseline for DRC-sensitive comparison until this interaction is better controlled. ## Split-Buffer Tier Selection Can Overload the Expensive Tier in Heterogeneous Designs ### Summary The old split-net tier selection policy was too narrow for heterogeneous upper/bottom technology stacks: - Cadence previously selected the split-buffer tier mainly by heavier fanout. - OpenROAD previously selected the split-buffer tier mainly by opposite-of-driver placement. Both policies can be wrong in heterogeneous designs. Heavy-fanout-only can push many buffers into the physically more expensive or already highly utilized tier, for example the 45nm side in an ASAP7/Nangate45 stack. Opposite-of-driver-only is more stable, but it ignores utilization stress and can still insert buffers into a tier that has little area margin. ### Current Fix Both Cadence and OpenROAD split-net implementations now use the same lightweight cost function: ```text score(t) = w_util * util_penalty(t) + w_hbt * hbt_penalty(t) + w_area * buffer_area_penalty(t) ``` Defaults: - `u_safe = 0.60` - `alpha = 12.0` - `w_util = 1.0` - `w_hbt = 2.5` - `w_area = 400.0` - high-util forbid threshold = `0.80` - near-tie threshold = `5%` The utilization term is a single-threshold exponential penalty, so utilization above `0.60` is penalized nonlinearly and `0.80+` becomes much more expensive than a linear fanout-only rule would imply. The HBT term uses a proxy named `estimated_extra_hbt`: - if `buffer_tier != driver_tier`, `estimated_extra_hbt = 1` - if `buffer_tier == driver_tier`, `estimated_extra_hbt = retained_opposite_tier_sink_count` This value is only a split-decision cost proxy. It is not the actual routed HBT count and should not be used as a final physical HBT metric. The area term uses the chosen split buffer master area normalized by the current core area: - `buffer_area_penalty = chosen_buffer_area / core_area` This intentionally makes the area term more visible on very small designs, where a few added buffers can perturb density much more strongly than on large designs. ### Remaining Limitations - Tier utilization is global and approximate; it is computed once per split run from tier-classified instance/master area divided by core area. - The new area term is still lightweight; it does not predict final routed congestion or final routed HBT count. It only scores the inserted split buffer area against the current core area. - The policy intentionally does not use local bin congestion, routing demand, or routed HBT estimates. - Buffer master selection and split verification remain governed by the existing flow rules. - This improves split-buffer tier selection quality, but it does not claim to predict final routed HBT count or final DRC behavior. ## OpenROAD ariane133 Has CTS and Pin-Access Blockers Historical note: `ariane133` has been removed from the `pesudo-3d` case collection. The observations below describe earlier runs and are retained for reference; they are not part of this branch's current regression set. ### Summary The OpenROAD 3D flow currently has unresolved `ariane133` blockers that should be tracked separately from the Cadence issues: - `nangate45_3D/ariane133` fails in `ord-cts` - `asap7_nangate45_3D/ariane133` fails in `ord-cts` - `asap7_3D/ariane133` can reach route but fails during GRT / detailed-router pin access These are treated as known OpenROAD flow blockers for the SRAM/macro-heavy `ariane133` design. ### Affected Scope - OpenROAD flow - `ariane133` - technologies: - `nangate45_3D` - `asap7_nangate45_3D` - `asap7_3D` ### Observed CTS Failure The 45nm and heterogeneous ariane133 OpenROAD runs fail at `ord-cts`. Observed logs: - [nangate45_3D ariane133 run log](run_logs/nangate45_3D/ord/run/ariane133_run.log) - `make: *** [Makefile:447: ord-cts] Error 11` - `clk_i_regs` has about `19940` register sinks - CTS creates about `3562` clock buffers / clock nets before failing - [asap7_nangate45_3D ariane133 run log](run_logs/asap7_nangate45_3D/ord/run/ariane133_run.log) - `make: *** [Makefile:447: ord-cts] Error 11` - `clk_i_regs` has about `19851` register sinks - CTS creates about `5122` clock buffers / clock nets before failing Current interpretation: - this is likely a CTS robustness / scaling / placement-legality issue in the OpenROAD 3D flow for ariane133 - it should not be diagnosed as a split-net-only issue without additional evidence - the 45nm and mixed-tech cases should not be used as valid final OpenROAD ariane133 results until this CTS blocker is fixed ### Observed GRT / Pin-Access Failure The 7nm ariane133 OpenROAD run can fail later during global route / detailed router pin access. Observed log: - [asap7_3D ariane133 GRT log](logs/asap7_3D/ariane133/openroad/5_1_grt.log) - `DRT-0073 No access point` - failures are reported on SRAM pins such as `rd_out[15]` - affected SRAM views include `sram_asap7_16x256_1rw_upper` and corresponding bottom-tier views - the stage exits at `Error: global_route.tcl, 103 DRT-0073` Current interpretation: - this points to a macro/SRAM LEF pin-access or route-access geometry issue for the ASAP7 ariane133 macro views - it is not expected to be solved by changing only the high-level global route command - the next debug target should be the SRAM macro pin shape / access-layer setup in the 7nm upper and bottom LEF views ### Current Status These OpenROAD ariane133 results should be treated as blocked: - `nangate45_3D/ariane133`: CTS blocker - `asap7_nangate45_3D/ariane133`: CTS blocker - `asap7_3D/ariane133`: GRT / DRT pin-access blocker Recommended next steps: - isolate CTS behavior on the 45nm and heterogeneous ariane133 runs - inspect SRAM macro pin-access geometry for the 7nm ariane133 run - keep these failures separate from Cadence placement/route issues and from the split-net cost-model tuning work