Default parameter extraction for simple configurations.

Free all device memory allocated with DeviceArray wrapper

Returns base arguments for kernel_tuner.tune_kernel().

Returns a list with all compiler options to pass to Kernel Tuner

Return the grid_div_x parameter for kernel_tuner.tune_kernel()

Default metrics for performance measurement.

Define constraints for what parameter combinations are valid during tuning. Two options: 1. Return list of string expressions that mu

Process a single specialization and extract its configuration. Args: spec_text: Text of the specialization to process.

Runs the default configuration if enabled.

Save tuning results and metadata to files.

Adds a single tuning execution for a given architecture

Adds a single benchmark execution for a given architecture

Adds a single file containing the results of benchmarks executed on a single architecture. The benchmarks within the file may belong

\brief Make sure visible operations are complete Ensure that following visible reads are not reordered before preceding atomic operations Similarly t

\brief Wait for all vector memory operations to complete This ensures that previous visible writes to vector memory have completed before the functio

Calculate and print performance statistics for each strategy. Returns the DataFrame for the heatmap.

\brief Computes the closest multiple of \p divisor to a certain \p ref. \param ref Number to be rounded up. \param divisor Number which closest multip

\brief Count leading zeroes Count the number of consecutive 0-bits, starting from the most significant bit. \p x must be non-zero.

combine()

Generate the content of the configuration file, including license and header guards, based on general template file.

Generate the content of the configuration file, including license and header guards, based on general template file.

Create a heatmap comparing different strategies' performance.

Filter algorithms based on regex pattern.

Generate wrapper code using Jinja2 template inheritance.

Return list of supported tuning algorithms.

Get the default configuration for a specific architecture and types. Args: arch: Target architecture (e.g., 'gfx1030').

Creates a hashable namedtuple based on selectors and instanced_types. Missing types are filled with EMPTY_TYPENAME.

Get the result from the benchmark json.

Retrieves the appropriate scoring function for the specified algorithm. This function ensures consistent configuration scoring between *

Reads selectors from a JSON file at selectors_path. The JSON should be a dict mapping algorithm names to lists of selector strings.

Returns the warp lane mask of all lanes less than or equal to the calling thread

Returns the warp lane mask of all lanes less than the calling thread