Placement Challenge

[sanjeevsinha1]

Summary:

• Implement loss functions for overlap and wirelength which have a smooth gradient
• Use KD-tree for sparse overlap loss computation (not across all (i, j) pairs) when design has lots of std cells. Maintains enough confidence where we are still dealing with valid overlaps between cells, but not doing a naive pairwise comparison (can be expensive) unless sparse overlap loss is considered ~0 as a final sanity check
• Different optimization configurations for larger designs to maintain stability between epochs and avoid noisy loss
• Include batched, vectorized computations for calculating overlap/cells with overlap due to memory constraints
• Full test suite passes with 0 overlap

Running:

python test.py

Results:

Average Overlap: 0.0000
Average Wirelength: 0.1493
Total Runtime: 4057.78s