Example: Maxwell Benchmark

Script: examples/benchmark_maxwell.py

Goal

Measure end-to-end runtime for tide.maxwelltm forward and forward+backward under fixed workloads.

Inputs

The script runs two scenarios:

• small: ny=96, nx=96, nt=128, n_shots=2
• medium: ny=160, nx=160, nt=192, n_shots=4

Common settings: stencil=2, pml_width=8, Ricker source. Precision is configurable with --compute-dtype and --mp-mode.

Steps

1. Pick runtime device in priority order: cuda, then cpu.
2. For each scenario and mode (forward, forward+backward):
3. Run 3 warmup iterations.
4. Run 10 timed iterations and synchronize device around each measurement.
5. Print mean/P50/P90 latency in milliseconds.

Output

CSV-like rows:

Case,Mode,Mean(ms),P50(ms),P90(ms)

Use these rows to compare before/after optimization on the same machine and config.