Maxwell Equation Simulator (CUDA-Accelerated)

This module accelerates a 2D Maxwell's Equations simulator that predicts electromagnetic wave propagation. The original CPU implementation took ~15 s for 4096^2 cells and ~4 min for 65,536^2 cells (~197,402,344 cells/s). The CUDA implementation achieves ~4,887,778,050 cells/s, exceeding the expected ~4,146,082,804 cells/s.

Key techniques:

• Port core field updates to the GPU with Thrust and device lambdas in maxwell.cu.
• Coarsen the grid by tiling and computing per-tile averages using cub::BlockReduce in coarse.cu for faster, grainier visualization.
• Assign one thread block per coarse tile; compute the average of thread values and write to the coarse grid.

Files:

• maxwell.cu: GPU implementation of field updates and the simulate loop.
• coarse.cu: Coarse-grid kernel using cub::BlockReduce to average tiles.

Assumptions:

• The shared support header dli.h provides types, constants like dli::C0, dimensions, and utilities used by these modules.

Build notes:

• Requires CUDA toolkit and CUB (modern CUDA ships CUB).
• Include paths must resolve dli.h and link with any project scaffolding using these functions.

Performance summary:

• Baseline CPU: ~197,402,344 cells/s.
• CUDA port: ~4,887,778,050 cells/s.
• With coarse tiling and cub::BlockReduce: ~4,902,636,317 cells/s.

Images

How the coarse.cu code works:

How the coarse.cu animation would look like: