CLI Reference

The primary entry point is gpu_main.py.

Usage

python gpu_main.py [OPTIONS]

Arguments

Argument	Type	Default	Choices	Description
--config	string	config.json	--	Path to configuration file
--mode	string	single	single, performance, monte_carlo	Simulation mode
--profile	flag	false	--	Enable TensorFlow profiling
--runner	string	full	full, chunked, no_save	Simulation runner mode
--chunk-size	int	100	--	Chunk size for chunked runner

Modes

single (default)

Runs a single atmospheric turbulence simulation:

python gpu_main.py --mode single

performance

Runs the performance test suite, benchmarking the simulator across varying configurations (r0 values, resolutions, layer counts):

python gpu_main.py --mode performance

See Performance Testing for details.

monte_carlo

Runs a Monte Carlo analysis with repeated simulations and statistical aggregation:

python gpu_main.py --mode monte_carlo

See Monte Carlo for details.

Runner Modes

The --runner flag controls how frames are stored during a single simulation:

full

All frames are kept in GPU memory and saved as a single frames.npy file. Fastest for short simulations but uses the most memory.

python gpu_main.py --runner full

chunked

Processes frames in chunks and streams them to disk. Uses double-buffered writes for overlapping computation and I/O. Produces a frames_chunk_manifest.json index file.

python gpu_main.py --runner chunked --chunk-size 50

no_save

Runs the simulation without saving any frames. Useful for throughput benchmarking and testing.

python gpu_main.py --runner no_save

Examples

# Basic run with defaults
python gpu_main.py

# Custom config with profiling
python gpu_main.py --config my_config.json --profile

# Chunked streaming with small chunks
python gpu_main.py --runner chunked --chunk-size 25

# Performance test suite
python gpu_main.py --mode performance

# Monte Carlo analysis
python gpu_main.py --mode monte_carlo

# Dataset generation (configured via config.json create_data_set flag)
python gpu_main.py

TensorFlow Profiling

When --profile is passed, TensorFlow traces are saved to the output directory under monitoring/tensorboard/. View them with:

tensorboard --logdir outputs/YYYY-MM-DD/HH/RUN_N/monitoring/tensorboard/