scope_profiler.plotting_scripts.roofline

Roofline analysis from LIKWID FLOP-rate and memory-bandwidth metrics.

LIKWID reports its derived metrics per participating hardware thread. A region’s point therefore sums those rates: a parallel region that sustains 20 GFLOP/s on each of four cores belongs at 80 GFLOP/s, not at their mean. Arithmetic intensity is then GFLOP/s / GB/s and is numerically identical to FLOP/byte.

Functions

Name	Description
collect_roofline_points	Derive one roofline point per selected LIKWID region and rank.
plot_roofline	Plot attained performance against arithmetic intensity for LIKWID regions.

collect_roofline_pointsfunction#

View source

def collect_roofline_points(profiling_data: ProfilingResults | Sequence[ProfilingResults], *, flops_metric: str | None = None, bandwidth_metric: str | None = None, ranks: list[int] | int | None = None, include: list[str] | str | None = None, exclude: list[str] | str | None = None) -> list[dict]

Derive one roofline point per selected LIKWID region and rank.

flops_metric and bandwidth_metric select exact LIKWID derived metric names. When omitted, the first names containing FLOP/s and bandwidth respectively are selected for each region. Supplying exact names is recommended when a group offers both single- and double-precision FLOP rates.

plot_rooflinefunction#

View source

def plot_roofline(profiling_data: ProfilingResults | Sequence[ProfilingResults], *, flops_metric: str | None = None, bandwidth_metric: str | None = None, peak_flops: float | None = None, peak_bandwidth: float | None = None, ranks: list[int] | int | None = None, include: list[str] | str | None = None, exclude: list[str] | str | None = None, filepath: str | None = None, show: bool = False, verbose: bool = True, cmap: str = DEFAULT_CMAP, data_filepath: str | Path | None = None, data_format: str = 'csv', backend: str = 'matplotlib') -> dict

Plot attained performance against arithmetic intensity for LIKWID regions.

peak_flops (GFLOP/s) and peak_bandwidth (GB/s) should be measured hardware ceilings for a true roofline. If either is absent, the largest selected observation supplies that ceiling and the plot labels it empirical; it is useful for comparing regions but must not be read as a machine limit.