BEOL/FEOL metal fill with recursive partitioning (C++)

[sethupathib]

Overview

New C++17 project in gpu_metal_fill/: a dependency-free dummy metal fill engine for FEOL base layers (OD, PO) and BEOL metals (M1..M14). It uses recursive quadtree partitioning (like commercial fill flows) so leaves can be filled independently/in parallel and then merged. It also ships a synthetic GPU-block GDS generator for realistic testing.

GPU offload is intentionally deferred: the compute-heavy stages are isolated behind a FillBackend interface (CPU/OpenMP now, CUDA later).

Pipeline (per layer)

1. Rasterize existing geometry into an occupancy grid.
2. Windowed density via a summed-area table (O(1) per window).
3. Keep-out halo = morphological dilation by the spacing rule.
4. Fill on a pitch lattice, per window, up to a per-window target, never exceeding max, never entering keep-out.
5. Iterative DRC: measure achieved density, raise targets for windows still below min or too far below a denser neighbor (gradient), refill until converged.

Recursive partitioning (correctness at boundaries)

• Lattice-aligned cuts at lcm(density_window, fill_pitch) so windows and fill positions are identical globally vs per-partition.
• Halo/guard band: each leaf reads neighboring geometry as read-only context but only emits fill in its disjoint core, so merge = concatenation.
• Unit test asserts partition == global (merged result identical to a single-shot fill).

Synthetic GPU block (make_gpu_block)

A structurally realistic (not real) GPU block: 3×3 SM tiles each with SRAM macros (register file + shared memory) and standard-cell logic, routing channels, clock/spine routes (M9/M10), and a sparse top-metal power grid (M11–M14). Each layer gets a distinct density profile.

Result on the generated block (180×180 µm, 15,885 polygons, 52 partitions/layer): 1.39M fill shapes, 0 DRC violations.

M1 (logic + SRAM macros) and M12 (power grid) — red = quadtree partition cores, navy = existing, orange = dummy fill (note keep-out halos around the power straps):

GPU block M1 before
GPU block M1 after
GPU block M12 power grid after

Full GPU-block report: gpu_fill_report.txt

Build / run

cd gpu_metal_fill
make            # tools -> build/
make test       # 77 checks, 0 failures
make demo       # simple synthetic layout
make gpu-demo   # realistic GPU-block layout

Notes

• The default layer map / rules in layermap.cpp are illustrative, not a real PDK.
• CUDA backend and via fill are follow-ups.

_{To show artifacts inline, enable in settings.}