Skip to content

README.md

Latest commit

 

History

History
42 lines (26 loc) · 979 Bytes

README.md

File metadata and controls

42 lines (26 loc) · 979 Bytes

cuda.bindings Benchmarks

Usage

Requires pixi.

There are a couple of environments defined based on how cuda.bindings is installed:

  • wheel: Installs from conda packages
  • source: Installs from source

There are a couple of tasks defined:

  • bench: Runs the Python benchmarks
  • bench-cpp: Runs the C++ benchmarks

To run the benchmarks combine the environment and task:

# Run the Python benchmarks in the wheel environment
pixi run -e wheel bench

# Run the Python benchmarks in the source environment
pixi run -e source bench

# Run the C++ benchmarks (environment is irrelavant here)
pixi run -e wheel bench-cpp

pyperf JSON

The benchmarks are run using pyperf. The results are written to a JSON file in the format expected by pyperf.

The C++ benchmarks also generate a valid JSON file, in the same format.

pixi run -e wheel bench-cpp -0 cpp.json

pixi run -e wheel pyperf stats cpp.json