Custom CUDA-backend based Optimization

[bala-git9]

Placement Challenge Submission - BALA

Overview

This submission implements a modular, optimized solution for the VLSI cell placement challenge, achieving significant performance improvements through custom CUDA acceleration, spatial hashing, and advanced loss function design.

Key Improvements

Performance:

• 200-267x speedup for test case 11 (10k cells) with custom CUDA backend
• 13.5-15x speedup for test case 12 (100k+ cells)
• Spatial hashing reduces complexity from O(N²) to O(N) for large problems

Technical Innovations:

• Custom CUDA backend for overlap computation (test cases 11-12)
• StrongFastSigmoid surrogate gradient function (inspired by SNN research)
• Adaptive hyperparameter scheduling based on problem size
• Margin-based overlap loss with conditional application
• Extended training (5000 epochs) for large problems

Code Organization:

• Modularized original monolithic placement.py into maintainable components
• Preserved original test structure and API compatibility
• Comprehensive documentation in CHANGES.md

Results

• Test cases 1-10: Zero overlap achieved
• Test case 11: Zero overlap, ~37.5s for 5000 epochs
• Test case 12: 0.0007 overlap ratio (71/100010 cells), ~541s for 5000 epochs

Files Changed

• Modularized codebase into placement_modules/ directory
• Added cuda_backend/ for CUDA-accelerated overlap computation
• Updated test.py to work with modular structure
• Added comprehensive documentation (CHANGES.md, PERFORMANCE_COMPARISON.md)

See CHANGES.md for detailed technical documentation.