tilelang.profiler.wall¶

Wall-clock benchmarking with optional explicit device synchronization.

Functions¶

bench_with_wall(fn, n_repeat[, quantiles, ...])

Time each call in milliseconds, including optional synchronization.

Module Contents¶

tilelang.profiler.wall.bench_with_wall(fn, n_repeat, quantiles=None, return_mode='mean', synchronize=None)¶

Time each call in milliseconds, including optional synchronization.

Synchronize before starting each sample to exclude preceding device work, then wait for the timed call to complete before stopping the clock. Quantiles use linear interpolation; median returns the lower middle sample, matching the GPU benchmark’s aggregation conventions.

Parameters:

  • fn (collections.abc.Callable)

  • n_repeat (int)

  • quantiles (list[float] | None)

  • return_mode (Literal['min', 'max', 'mean', 'median'])

  • synchronize (collections.abc.Callable | None)

Return type:

float | list[float]