Repository navigation
Conversation
The upstream GT2N LEFs carry no threshold-voltage information, so the
resizer, which identifies a cell's Vt from IMPLANT-layer obstructions,
sees every W/Vt flavor as the same Vt. vt_swap never fires (every
repair_timing reports 0 VT), keep_sizing_vt has no effect, and Vt only
changes as a side effect of sizeup.
Declare ELVT/ULVT/LVT/SVT/HVT IMPLANT layers in gt2_tech.lef and give
every macro a full-cell OBS rectangle on its Vt layer, as asap7 does.
The rectangles mirror the GDS, where each elvt/ulvt/svt/hvt cell has a
single Vt-layer rectangle from the origin to the cell SIZE. LVT is the
base device with no GDS layer; it gets an LEF-only layer so every
flavor is explicit. LEF obstructions are not streamed out, so the
merged GDS is unchanged.
add_vt_implant.py makes the edit and is idempotent; rerun it after
re-syncing the LEFs from upstream.
gcd finish, before -> after:
yosys flow: WNS -68.8 -> -60.0 ps, TNS -1208 -> -568 ps,
leakage 43.6 -> 33.8 uW
syn flow: WNS -216.0 -> -45.1 ps, TNS -8042 -> -794 ps
Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
There was a problem hiding this comment.
Code Review
This pull request introduces a Python script, add_vt_implant.py, and updates the GT2N PDK LEF files to declare and add threshold-voltage (Vt) implant layers (ELVT, ULVT, LVT, SVT, HVT) as obstructions in the cell macros. This enables OpenROAD's resizer to identify a cell's Vt and perform Vt-aware optimizations. The review feedback recommends specifying encoding='utf-8' for all file read and write operations in the script to prevent platform-dependent encoding issues, and adding a warning check in the main function in case no macro LEF files are found.
| def add_macro_obs(path, layer): | ||
| lines = path.read_text().splitlines(keepends=True) |
There was a problem hiding this comment.
Specify encoding="utf-8" when reading the macro LEF files to prevent platform-dependent encoding issues.
| def add_macro_obs(path, layer): | |
| lines = path.read_text().splitlines(keepends=True) | |
| def add_macro_obs(path, layer): | |
| lines = path.read_text(encoding="utf-8").splitlines(keepends=True) |
🔍 QoR checkMetrics reflect the PR merge build — i.e. what will land on the target branch. Commit 62 design(s) checked — 0 with regression(s), 0 without a comparable baseline. |
Summary
The upstream GT2N LEFs carry no threshold-voltage information. OpenROAD's resizer identifies a cell's Vt from IMPLANT-layer obstructions in the LEF (
Resizer::cellVTType), so on gt2n it sees every W/Vt flavor as the same Vt:vt_swapnever fires — everyrepair_timingon gt2n reports0 VTkeep_sizing_vthas no effectsizeuptreating Vt variants as drive strengthsThis adds implant layers the same way asap7 does:
gt2_tech.lef:ELVT/ULVT/LVT/SVT/HVTTYPE IMPLANTlayers after the front-end MASTERSLICE layers.gt2_6t_w*_*.leffiles gets a full-cellOBSrectangle on its Vt layer (creating theOBSblock for the filler and backside tap, which had none).gt2.layermap) from the origin to the cellSIZE(checked on all 720 cells). LVT is the base device with no GDS layer; it gets an LEF-only layer so every flavor is explicit and reports asLrather than-. LEF obstructions are not streamed out, so the merged GDS is unchanged.add_vt_implant.pymakes the edit and is idempotent; the README notes it as a local modification to rerun after re-syncing the LEFs from upstream.With this,
report_equiv_cells -vtreports H/S/L/UL/EL, andrepair_timingperforms Vt swaps (e.g.228 resized: 12 up, 216 VT).gcd finish QoR, before → after
SYNTH_USE_SYN=1The large syn improvement is because
repair_design -pre_placementin the syn flow flips the mapped netlist to nearly all HVT (resizeToCapRatiobreaks equal-area ties toward the weakest drive); with Vts now recognized,repair_timingcan swap those back. MakingresizeToCapRatioVt-preserving is a separate OpenROAD change.