AGENT SKILL

macro_pin_side

暂无简介

License 许可
Apache-2.0
所属仓库
FeSens/openTPU
最近更新
2026-11-12
Stars
★ 971
Forks
96

02 资源文件 1

03 SKILL.md 正文

Skill: Macro pin-side highlight

Detect the signal-pin side(s) of every macro from OpenROAD DB geometry and draw a thick highlight line on every pin-bearing edge in a whole-die PNG. A macro may have pins on more than one side (e.g. an SRAM with pins on both L and R) — each real pin side is highlighted, not just the single dominant one.

Key idea: never infer the pin side visually from a screenshot. The side comes from iterm getBBox, which already reflects the macro's current placed orientation (R0/R90/MX/MY…). The screenshot is only the canvas we annotate.

What it does

  1. OpenROAD loads the placement (--odb, or --def + LEFs).

  2. For every BLOCK instance:

    • take the placed bbox (inst getBBox),
    • iterate inst getITerms, keep only getSigType == SIGNAL,
    • compute each pin center from iterm getBBox,
    • measure its distance to the macro L/R/T/B edges and count it toward the nearest edge,
    • dominant side = edge with the most pins (NONE if no signal pins).
  3. Writes structured text, one line per macro (the per-edge counts are the source of truth; pin side(s) are derived from them in Python — see step 5):

    MACRO_PIN_SIDE name=<inst> orient=<orient> side=<L/R/T/B/NONE> dominant=<n> \
        counts=L:<n>,R:<n>,T:<n>,B:<n> bbox=(x0,y0,x1,y1)
    

    (bbox in microns. side = the single dominant edge; counts carry all four so a multi-side macro can be recovered without re-running OpenROAD.)

  4. save_image of the whole die, following the whole-die-with-pins convention of skills_Nangate45/image_snapshot: the save area is the union of the die box and the top-level IO-pin (bterm) boxes, padded by margin, and the signal/clock bterm boxes are recorded so they can be redrawn (OpenROAD rasterizes thin pins to near-invisibility at die zoom).

  5. Python derives each macro's pin side(s) from the per-edge counts with pin_sides(counts, min_pins): an edge is a pin side when it carries at least --min-pins signal pins (default 1 = every edge with any pins). Each pin is already binned to its geometrically nearest edge, so this reports all real pin sides — a macro with address/data on L+R and control on T+B shows four bars (e.g. L:40 R:38 T:10 B:10 → L,R,T,B). Then it maps each macro/pin bbox micron→pixel (with Y-flip) and draws:

    • the IO pins as small cyan boxes (so the die-edge pins stay visible), then
    • a black-shadow + bright-red bar on each pin side of every macro, drawn just inside that macro, hugging the pin edge and inset to ~70% of the edge length. Keeping it inside (not in the channel) means the bar is never confused with the neighbouring macro across an abutting edge,
    • a numbered label (yellow-on-black) drawn once per macro, on the dominant pin side. The number is the macro's index in the printed summary / MACRO_PIN_SIDE order, so you can read bar → macro name. Disable with --no-label.

    The stdout summary prints sides=<comma list> and a side_combos histogram (e.g. {'LR': 12}) so multi-side macros are obvious without opening the image.

Usage

conda activate coder

python skills/skills_Nangate45/macro_pin_side/macro_pin_side.py \
    --def     Benchmark/PlaceReasoner-Bench/bp_fe/def/u65_placed.def \
    --lef-dir Benchmark/PlaceReasoner-Bench/bp_fe/lef \
    --out-dir /…/OpenLayout_macro/work_images \
    --design  bp_fe_u65

From a checkpoint ODB instead (one read_db restores tech + LEF + placement):

python macro_pin_side.py --odb run/floorplan.odb --out-dir work_images --design foo

Options

  • --lef LEF1 LEF2 … — explicit LEFs (tech first) instead of --lef-dir glob. --lef-dir globs *.lef and reads any name containing tech first.
  • --width (default 1600) — save_image pixel width; height keeps aspect.
  • --margin (default 0.03) — pad fraction on the die∪pins union, pad = max(w,h) * margin.
  • --min-pins (default 1) — an edge is highlighted as a pin side when it carries at least this many signal pins. Default 1 shows every edge that has any pins (a four-side macro gets four bars). Raise it (e.g. 3) to suppress edges with only a stray pin or two.
  • --no-pins — skip the IO-pin overlay (pins on by default, matching image_snapshot).
  • --no-label — skip the numbered macro labels next to each bar.
  • --openroad (default /usr/bin/openroad).

Outputs (in --out-dir, prefixed <design>_)

  • <design>_pin_side.txt — the MACRO_PIN_SIDE lines (machine-readable).
  • <design>_iopins.txt — IO_PIN sig=… box=(…) lines for the bterm overlay.
  • <design>_pin_side.png — annotated whole-die-with-pins image.
  • <design>_raw.png — un-annotated OpenROAD render.
  • <design>_pin_side.tcl — the generated TCL (re-runnable / inspectable).

The driver also prints a per-macro summary and the side distribution.

Notes / gotchas

  • Headless render needs xvfb-run in this repo (offscreen/xcb both fail — see memory feedback_openroad_save_image). The driver already wraps OpenROAD in xvfb-run -a.
  • parse_sides() and draw_overlay() are importable if you want to reuse the highlight on an image you rendered elsewhere — pass the same micron area that save_image -area used (the driver reads it from the IMAGE_AREA stdout line).
  • Verified on bp_fe u65_placed.def: 11 BLOCK macros, 8×R / 3×L; red line lands on the correct geometric edge (right edge present, left edge clean) under the Y-flip.

04 同仓库更多 Skills