kernelfoundry.algorithm.utils.score

Functions

combine_gpu_arch_results(exec_result_dict, ...)	Combine results from different gpu archs by taking the worst correctness and average runtime.
combine_template_results(exec_result_list, ...)	Take the results for each parameter option and combine them into one result
compute_runtime_improvement(exec_result, ...)	Add improvement of kernel over baseline time to exec result, using ORM queries.
fastp(is_correct, baseline_speed, ...)	Rate of samples within a threshold p
geometric_mean_speed_ratio_correct_and_faster_only(...)	Geometric mean of the speed ratio for correct samples that have speedup > 1
geometric_mean_speed_ratio_correct_only(...)	Geometric mean of the speed ratio for correct samples
get_timing_stats(elapsed_times[, device, ...])	Get timing statistics from a list of elapsed times.
select_best_solution(eval_results)	Select the best kernel based on composite fitness score.

kernelfoundry.algorithm.utils.score.compute_runtime_improvement(exec_result, level: int, task_name: str, gpu_arch: str)

Add improvement of kernel over baseline time to exec result, using ORM queries.

Parameters:

• exec_result – Execution result containing runtime but missing runtime_improvement

• level – Level of the operation (e.g., 1)

• task_name – Operation name (with .py!)

• gpu_arch – gpu arch string

• db – Database object with SessionRO() context manager

kernelfoundry.algorithm.utils.score.select_best_solution(eval_results: list[EvalResult]) → int

Select the best kernel based on composite fitness score.

Uses EvalResult.compute_performance_score to ensure consistent scoring across the system.

Parameters:

eval_results – List of EvalResult

Returns:

index of best solution

Return type:

int

kernelfoundry.algorithm.utils.score.combine_gpu_arch_results(exec_result_dict: dict[str, EvalResult], combined_eval_log: str) → EvalResult

Combine results from different gpu archs by taking the worst correctness and average runtime.

kernelfoundry.algorithm.utils.score.combine_template_results(exec_result_list: list, parameter_options: list, return_index: bool = False)

Take the results for each parameter option and combine them into one result

kernelfoundry.algorithm.utils.score.geometric_mean_speed_ratio_correct_only(is_correct: ndarray, baseline_speed: ndarray, actual_speed: ndarray, n: int) → float

Geometric mean of the speed ratio for correct samples

kernelfoundry.algorithm.utils.score.geometric_mean_speed_ratio_correct_and_faster_only(is_correct: ndarray, baseline_speed: ndarray, actual_speed: ndarray, n: int) → float

Geometric mean of the speed ratio for correct samples that have speedup > 1

kernelfoundry.algorithm.utils.score.fastp(is_correct: ndarray, baseline_speed: ndarray, actual_speed: ndarray, n: int, p: float) → float

Rate of samples within a threshold p

kernelfoundry.algorithm.utils.score.get_timing_stats(elapsed_times: list[float], device=None, device_name: str = 'unknown') → dict

Get timing statistics from a list of elapsed times.

Parameters:

• elapsed_times – List of elapsed times in milliseconds

• device – CUDA device, record device info

Returns:

Dict containing mean, std, min, max and num_trials all timing are in ms