#!/usr/bin/env python3 # Copyright (c) 2025-2026 Advanced Micro Devices, Inc. All rights reserved. # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. """ This Python script is intended for the creation of autotuned configurations for the supported rocPRIM algorithms based on tuning results. The script does not update the configurations automatically, the user is responsible for installation and the correctness of the files """ import json import re import argparse import os import sys import collections import copy from dataclasses import dataclass from collections import defaultdict from typing import Dict, List, Callable, Optional, Tuple from jinja2 import Environment, PackageLoader TARGET_ARCHITECTURES = [ "gfx803", "gfx900", "gfx906", "gfx908", "gfx90a", "gfx942", "gfx1030", "gfx1100", "gfx1102", ] # C++ typename used for optional types EMPTY_TYPENAME = "empty_type" env = Environment( loader=PackageLoader("configs_header_generation"), lstrip_blocks=True, trim_blocks=True, ) class NotSupportedError(Exception): """Exception raised for algorithms that are not supported Attributes: message -- explanation of the error """ def __init__(self, message): self.message = message super().__init__(self.message) @dataclass class SelectionType: """ Data class describing a type used to select a configuration. """ name: str # True if rocprim::empty_type is a valid type in the algorithm's configuration. is_optional: bool # True if only the size of the type is relevant. If false, selection also happens based on rocprim::is_floating. # Only selecting on size will limit the number of tunings required and configs generated. # For example, a sort by key will not perform any logic on the value type and only move the data. # Hence it can select the value type on size only. select_on_size_only: bool @dataclass class SelectionConst: """ Data class describing a constant (integral, enum) used to select a configuration. """ name: str @dataclass class FallbackCase: """ Data class describing a fallback case: a description for a set of types that use the same tuned configuration. """ # Optional regex filter for the algorithm name. algo_regex: Optional[str] # C++ type for which to take the best tuned configuration. based_on_type: str # Inclusive maximum size of the type. sizeof_max: int # Exclusive minimum size of the type. sizeof_min: int # Whether the type has a floating point. is_floating_point: bool def translate_settings_to_cpp_metaprogramming( fallback_configuration: List[Tuple[SelectionType, FallbackCase]], const_configuration: List[SelectionConst], ) -> str: """ Translates a list of named fallback configuration entries to C++ metaprogramming idioms. """ setting_list: List[str] = [] for config_selection_type, fallback_entry in fallback_configuration: typename: str = config_selection_type.name if fallback_entry.based_on_type == EMPTY_TYPENAME: # If the entry is based on the empty type # (which is not present in the fallback file, but separately inserted) setting_list.append( f"(std::is_same<{typename}, rocprim::{EMPTY_TYPENAME}>::value)" ) else: # Only add a floating-point check on the first selection type. For the remaining selection types, a limited # number of fallbacks are generated, which are based on the integral types. if not config_selection_type.select_on_size_only: negation: str = "" if fallback_entry.is_floating_point else "!" output: str = ( negation + f"bool(rocprim::is_floating_point<{typename}>::value)" ) setting_list.append(output) setting_list.append(f"(sizeof({typename}) <= {fallback_entry.sizeof_max})") # sizeof(type) will always be greater than zero, can omit the check if min is zero or smaller if fallback_entry.sizeof_min > 0: setting_list.append( f"(sizeof({typename}) > {fallback_entry.sizeof_min})" ) # If the fallback entry has a sizeof of one (only true for int8_t) and the associated type is optional, # we need an additional check since the empty type also has a sizeof of one if ( fallback_entry.based_on_type == "int8_t" and config_selection_type.is_optional ): setting_list.append( f"(!std::is_same<{typename}, rocprim::{EMPTY_TYPENAME}>::value)" ) for name, value in const_configuration.items(): setting_list.append(f"({name} == {value})") return "std::enable_if_t<(" + " && ".join(setting_list) + ")>" class TuningsOfArchitecture: """ Stores the tuning results for a specific architecture and algorithm. """ def __init__( self, arch_name: str, config_selection_params, fallback_entries: List[FallbackCase], config_get_best, algorithm_name, ): self.config_selection_params = config_selection_params self.fallback_entries: List[FallbackCase] = fallback_entries self.arch_name: str = arch_name self.config_get_best: Callable[[Dict], Dict[str, str]] = config_get_best self.algorithm_name: str = algorithm_name # Dictionary storing the tunings # Key is an instantiation of the configuration selection types # Value is a list of all tuning runs corresponding to that instantiation, # these tunings in this list vary in the actual configuration used to run the tuning self.tunings = defaultdict(list) def __get_instance_key(self, instanced_types): """ Takes in a list of instantiated types in the form of (name, value)-pairs for some 'name' in the selection types. Returns a named tuple where None values are converted to empty_type. """ Instance = collections.namedtuple( typename="Instance", field_names=[cfg_param.name for cfg_param in self.config_selection_params], ) # Convert None values and "None" string to empty_type values = { field: ( EMPTY_TYPENAME if instanced_types[field] is None or instanced_types[field] == "None" else instanced_types[field] ) for field in Instance._fields } return Instance(**values) def add_measurement(self, cache_data: Dict[str, str]): """ Adds a single tuning run. """ instance_key = self.__get_instance_key(cache_data) self.tunings[instance_key] = list(cache_data["cache"].values()) @property def name(self) -> str: return self.arch_name def __get_best_tuning(self, instance_key) -> Dict[str, str]: """ Returns the best performing tuning from a list of tunings. For now, use the items per second as metric. In case the tuning with the given configuration is not present, None is returned. """ if instance_key in self.tunings.keys(): return self.config_get_best(self.tunings[instance_key]) else: return None @property def best_config_by_selection_types(self): """ Returns a dictionary containing each instantion of the selection configuration as a key and the single best performing tuning run as a value. """ output = {} for instance, tunings in self.tunings.items(): output[instance] = self.__get_best_tuning(instance) return output def __add_fallback_to_output( self, output, fallback_configuration: List[Tuple[SelectionType, FallbackCase]], const_configurations: List[SelectionConst], ): """ Searches for specific fallback configuration in tuning results and adds it to the output :param output: list of strings to append to :param fallback_configuration: the configuration to look for :param const_configurations: list of constant selection parameters :return: """ search_key: Dict[str, str] = { config_selection_type.name: fallback_entry.based_on_type for (config_selection_type, fallback_entry) in fallback_configuration } for const_configuration in const_configurations or [{}]: for k, v in const_configuration.items(): search_key[k] = v best_tuning_result: Dict[str, str] = self.__get_best_tuning( self.__get_instance_key(search_key) ) print_config: str = ", ".join([f"{k} = {v}" for k, v in search_key.items()]) if best_tuning_result is None: print( f"WARNING {self.name}: No {self.algorithm_name} measurement found for creating fallback configuration " f'entry for "{print_config}"' ) else: output.append( ( print_config, translate_settings_to_cpp_metaprogramming( fallback_configuration, const_configuration ), best_tuning_result, ) ) @property def fallback_types(self): """ Provides a fallback triplet of (string describing the type used for generating the fallback, C++ enable if statement, tuning containing the selected parameters for the algorithm). This function only supports algorithms with at most two types. """ output = [] # Collect all combination of constant selection parameters from avaliable tuning results const_configurations = [] config_selection_const_names = [ cfg_param.name for cfg_param in self.config_selection_params if isinstance(cfg_param, SelectionConst) ] if config_selection_const_names: Consts = collections.namedtuple( typename="Consts", field_names=config_selection_const_names ) all_consts = [ Consts( **{ k: v for k, v in instance_key._asdict().items() if k in config_selection_const_names } ) for instance_key in self.tunings.keys() ] const_configurations = [c._asdict() for c in sorted(set(all_consts))] config_selection_types = [ cfg_param for cfg_param in self.config_selection_params if isinstance(cfg_param, SelectionType) ] # If there are more than two selection types, do not generate fallback cases # Otherwise, too many tunings would need support for the full product of fallback entries if len(config_selection_types) > 2: print( f"INFO: not generating fallbacks for {self.algorithm_name} as it has too many types." ) return output # If the type requires selection on size only, the floating-point fallback entries are skipped. # The decision to skip floating-point entries instead of integral ones is arbitrary. def skip_entry( config_selection_type: SelectionType, fallback_entry: FallbackCase ) -> bool: return ( config_selection_type.select_on_size_only and fallback_entry.is_floating_point ) # If there is exactly one type, generate full fallbacks. Note that this type will never be optional if len(config_selection_types) == 1: config_selection_type = config_selection_types[0] if config_selection_type.is_optional: raise ( ValueError( f'Algorithm "{self.algorithm_name}" has a single type that is optional' ) ) for entry in self.fallback_entries: if skip_entry(config_selection_type, entry): continue # Let the single selection type be based on the current fallback entry fallback_configuration = [(config_selection_type, entry)] # Find the closest measurement and create the config line self.__add_fallback_to_output( output, fallback_configuration, const_configurations ) # If there are two types, generate full fallbacks for the first type, but limited fallbacks for the second type if len(config_selection_types) == 2: # Assume that the first type is not optional if config_selection_types[0].is_optional: raise ( ValueError( f'Algorithm "{self.algorithm_name}" has two types, but the first is optional' ) ) # Enforce that only one type selects both on size and number representation (integral or floating) # to limit the number of generated configs. if ( not config_selection_types[0].select_on_size_only and not config_selection_types[1].select_on_size_only ): raise ( ValueError( f'Algorithm "{self.algorithm_name}" has two types, but neither only selects on size' ) ) # Only based_on_type is relevant empty_fallback = FallbackCase(None, EMPTY_TYPENAME, 0, 0, False) # If a type is optional, also generate the fallbacks where the type is empty. fallback_entries_0: List[FallbackCase] = self.fallback_entries.copy() if config_selection_types[0].is_optional: fallback_entries_0.append(empty_fallback) fallback_entries_1: List[FallbackCase] = self.fallback_entries.copy() if config_selection_types[1].is_optional: fallback_entries_1.append(empty_fallback) fallback_configuration: List[Tuple[SelectionType, FallbackCase]] = [] for fallback_0 in fallback_entries_0: if skip_entry(config_selection_types[0], fallback_0): continue fallback_configuration.append((config_selection_types[0], fallback_0)) for fallback_1 in fallback_entries_1: if skip_entry(config_selection_types[1], fallback_1): continue fallback_configuration.append( (config_selection_types[1], fallback_1) ) self.__add_fallback_to_output( output, fallback_configuration, const_configurations ) fallback_configuration.pop() fallback_configuration.pop() return output # Default formula to pick the best configuration, only looking at time. def default_config_get_best(input: Dict) -> Dict[str, str]: return min( input, key=lambda x: ( float("inf") if isinstance(x.get("time"), str) else x.get("time", float("inf")) ), ) # If we can double the sorted items_per_block, and items_per_second does not degrade more than ~10%, consider it superior. def block_sort_config_get_best(input: Dict) -> Dict[str, str]: return min( input, key=lambda x: x.get("time", float("inf")) * ((float(x["cfg"]["bs"]) * float(x["cfg"]["ipt"])) ** (1 / 4)), ) # Best configuration is a combination between best oddeven and best mergepath impl. # We use oddeven only for small input sizes (< ~200K), so it is a hardcoded value which is the best for almost all cases. # You can find this value in the tuning template def merge_sort_block_merge_config_get_best(input: Dict) -> Dict[str, str]: input_mergepath = list( filter(lambda x: (int(x.get("cfg").get("oddeven_size_limit")) == 0), input) ) # Since merge_sort_block_merge is used after radix_sort_block_sort<256, 4>, and # mergepath_block_size * mergepath_items_per_thread >= 256*4 should hold (TODO: this will be solved in the near future): input_mergepath = list( filter( lambda x: ( int(x.get("cfg").get("mergepath_bs")) * int(x.get("cfg").get("mergepath_ipt")) <= 1024 ), input_mergepath, ) ) best_mergepath = min(input_mergepath, key=lambda x: x.get("time", float("inf"))) return best_mergepath class Algorithm: """ Aggregates the data for an algorithm, including the generation of the configuration file. """ def __init__( self, fallback_entries: List[FallbackCase], config_get_best=default_config_get_best, ): self.architectures: Dict[str, TuningsOfArchitecture] = {} self.fallback_entries: List[FallbackCase] = fallback_entries self.config_get_best = config_get_best def add_measurement(self, single_tuning_data: Dict[str, str], architecture: str): """ Adds a single tuning execution for a given architecture """ if architecture not in self.architectures: self.architectures[architecture] = TuningsOfArchitecture( architecture, self.config_selection_params, self.fallback_entries, self.config_get_best, self.algorithm_name, ) self.architectures[architecture].add_measurement(single_tuning_data) def create_config_file_content(self) -> str: """ Generate the content of the configuration file, including license and header guards, based on general template file. """ if "target_arch::gfx908" in self.architectures: self.architectures["target_arch::unknown"] = copy.deepcopy( self.architectures["target_arch::gfx908"] ) self.architectures["target_arch::unknown"].arch_name = ( "target_arch::unknown" ) algorithm_template = env.get_template(self.cpp_configuration_template_name) rendered_template = algorithm_template.render( all_architectures=self.architectures.values() ) return rendered_template """ Each algorithm uses ninja templates to generate C++ configuration specification. The generated configuration file contains configs for two cases: - No architecture or instantiation of configuration selection types is provided (general base case). - The architecture and configuration selection types are provided, the configuration is based on the tuning results of similar types (fallback case). config_selection_params is a list of parameters (types, constants) that are used to select a configuration. The fallback file will be used to generate the fallback cases. If there are two types, one of these will only select on the size of the type (defined by the config selection parameters), limiting the number of fallback cases. If the type is optional, additional fallback configurations will be generated that match the case when the optional selection type passed by the user is rocprim::empty_type. The config_selection_params should specify at least one non-optional type. The optional type should not be the first type. """ class AlgorithmDeviceHistogram(Algorithm): algorithm_name = "device_histogram" cpp_configuration_template_name = "histogram_config_template" config_selection_params = [ SelectionType(name="value_type", is_optional=False, select_on_size_only=False), SelectionConst(name="channels"), SelectionConst(name="active_channels"), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceMergeSortBlockSort(Algorithm): algorithm_name = "device_merge_sort_block_sort" cpp_configuration_template_name = "mergesort_block_sort_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False), SelectionType(name="value_type", is_optional=True, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries, block_sort_config_get_best) class AlgorithmDeviceMergeSortBlockMerge(Algorithm): algorithm_name = "device_merge_sort_block_merge" cpp_configuration_template_name = "mergesort_block_merge_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False), SelectionType(name="value_type", is_optional=True, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__( self, fallback_entries, merge_sort_block_merge_config_get_best ) class AlgorithmDeviceRadixSortBlockSort(Algorithm): algorithm_name = "device_radix_sort_block_sort" cpp_configuration_template_name = "radixsort_block_sort_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False), SelectionType(name="value_type", is_optional=True, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries, block_sort_config_get_best) class AlgorithmDeviceRadixSortOnesweep(Algorithm): algorithm_name = "device_radix_sort_onesweep" cpp_configuration_template_name = "radixsort_onesweep_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False), SelectionType(name="value_type", is_optional=True, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceReduce(Algorithm): algorithm_name = "device_reduce" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False) ] cpp_configuration_template_name = "reduce_config_template" def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceScan(Algorithm): algorithm_name = "device_scan" cpp_configuration_template_name = "scan_config_template" config_selection_params = [ SelectionType(name="value_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) # TODO: select_on_size_only may need to get re-evaluated, as it is likely that swapping the value # of select_on_size_only for key and value gives better results class AlgorithmDeviceScanByKey(Algorithm): algorithm_name = "device_scan_by_key" cpp_configuration_template_name = "scanbykey_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False), SelectionType(name="value_type", is_optional=False, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceBinarySearch(Algorithm): algorithm_name = "device_binary_search" cpp_configuration_template_name = "binary_search_config_template" config_selection_params = [ SelectionType(name="value_type", is_optional=False, select_on_size_only=False), SelectionType(name="output_type", is_optional=False, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceUpperBound(Algorithm): algorithm_name = "device_upper_bound" cpp_configuration_template_name = "upper_bound_config_template" config_selection_params = [ SelectionType(name="value_type", is_optional=False, select_on_size_only=False), SelectionType(name="output_type", is_optional=False, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceLowerBound(Algorithm): algorithm_name = "device_lower_bound" cpp_configuration_template_name = "lower_bound_config_template" config_selection_params = [ SelectionType(name="value_type", is_optional=False, select_on_size_only=False), SelectionType(name="output_type", is_optional=False, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceAdjacentDifference(Algorithm): algorithm_name = "device_adjacent_difference" cpp_configuration_template_name = "adjacent_difference_config_template" config_selection_params = [ SelectionType(name="value_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceAdjacentDifferenceInplace(Algorithm): algorithm_name = "device_adjacent_difference_inplace" cpp_configuration_template_name = "adjacent_difference_inplace_config_template" config_selection_params = [ SelectionType(name="value_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceAdjacentFind(Algorithm): algorithm_name = "device_adjacent_find" cpp_configuration_template_name = "adjacent_find_config_template" config_selection_params = [ SelectionType(name="input_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceSegmentedRadixSort(Algorithm): algorithm_name = "device_segmented_radix_sort" cpp_configuration_template_name = "segmented_radix_sort_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False), SelectionType(name="value_type", is_optional=True, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceSegmentedTopkAir(Algorithm): algorithm_name = "device_segmented_topk_air" cpp_configuration_template_name = "segmented_topk_air_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False), SelectionType(name="value_type", is_optional=True, select_on_size_only=True)] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceTransform(Algorithm): algorithm_name = "device_transform" cpp_configuration_template_name = "transform_config_template" config_selection_params = [ SelectionType(name="value_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceTopkAir(Algorithm): algorithm_name = "device_topk_air" cpp_configuration_template_name = "topk_air_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False), SelectionType(name="value_type", is_optional=True, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDevicePartitionTwoWayPredicate(Algorithm): algorithm_name = "device_partition_two_way_predicate" cpp_configuration_template_name = "partition_two_way_predicate_config_template" config_selection_params = [ SelectionType(name="data_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDevicePartitionTwoWayFlag(Algorithm): algorithm_name = "device_partition_two_way_flag" cpp_configuration_template_name = "partition_two_way_flag_config_template" config_selection_params = [ SelectionType(name="data_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDevicePartitionFlag(Algorithm): algorithm_name = "device_partition_flag" cpp_configuration_template_name = "partition_flag_config_template" config_selection_params = [ SelectionType(name="data_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDevicePartitionPredicate(Algorithm): algorithm_name = "device_partition_predicate" cpp_configuration_template_name = "partition_predicate_config_template" config_selection_params = [ SelectionType(name="data_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDevicePartitionThreeWay(Algorithm): algorithm_name = "device_partition_three_way" cpp_configuration_template_name = "partition_three_way_config_template" config_selection_params = [ SelectionType(name="data_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceSelectFlag(Algorithm): algorithm_name = "device_select_flag" cpp_configuration_template_name = "select_flag_config_template" config_selection_params = [ SelectionType(name="data_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceSelectPredicate(Algorithm): algorithm_name = "device_select_predicate" cpp_configuration_template_name = "select_predicate_config_template" config_selection_params = [ SelectionType(name="data_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceSelectPredicatedFlag(Algorithm): algorithm_name = "device_select_predicated_flag" cpp_configuration_template_name = "select_predicated_flag_config_template" config_selection_params = [ SelectionType(name="data_type", is_optional=False, select_on_size_only=False), SelectionType(name="flag_type", is_optional=False, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceSelectUnique(Algorithm): algorithm_name = "device_select_unique" cpp_configuration_template_name = "select_unique_config_template" config_selection_params = [ SelectionType(name="data_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceSelectUniqueByKey(Algorithm): algorithm_name = "device_select_unique_by_key" cpp_configuration_template_name = "select_unique_by_key_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False), SelectionType(name="value_type", is_optional=False, select_on_size_only=True), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceReduceByKey(Algorithm): algorithm_name = "device_reduce_by_key" cpp_configuration_template_name = "reduce_by_key_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=True), SelectionType(name="value_type", is_optional=False, select_on_size_only=False), ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceFindFirstOf(Algorithm): algorithm_name = "device_find_first_of" cpp_configuration_template_name = "find_first_of_config_template" config_selection_params = [ SelectionType(name="value_type", is_optional=False, select_on_size_only=True) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceRunLengthEncode(Algorithm): algorithm_name = "device_run_length_encode" cpp_configuration_template_name = "run_length_encode_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceRunLengthEncodeNonTrivial(Algorithm): algorithm_name = "device_run_length_encode_non_trivial" cpp_configuration_template_name = ( "run_length_encode_non_trivial_runs_config_template" ) config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False) ] def __init__(self, fallback_entries): Algorithm.__init__(self, fallback_entries) class AlgorithmDeviceMerge(Algorithm): algorithm_name = "device_merge" cpp_configuration_template_name = "merge_config_template" config_selection_params = [ SelectionType(name="key_type", is_optional=False, select_on_size_only=False), SelectionType(name="value_type", is_optional=True, select_on_size_only=True), ] def filt_algo_regex(e: FallbackCase, algorithm_name): if e.algo_regex: return re.match(e.algo_regex, algorithm_name) is not None return True def create_algorithm(algorithm_name: str, fallback_entries: List[FallbackCase]): fallback_entries = list( filter(lambda e: filt_algo_regex(e, algorithm_name), fallback_entries) ) if algorithm_name == "device_histogram": return AlgorithmDeviceHistogram(fallback_entries) elif algorithm_name == "device_merge_sort_block_sort": return AlgorithmDeviceMergeSortBlockSort(fallback_entries) elif algorithm_name == "device_merge_sort_block_merge": return AlgorithmDeviceMergeSortBlockMerge(fallback_entries) elif algorithm_name == "device_radix_sort_block_sort": return AlgorithmDeviceRadixSortBlockSort(fallback_entries) elif algorithm_name == "device_radix_sort_onesweep": return AlgorithmDeviceRadixSortOnesweep(fallback_entries) elif algorithm_name == "device_reduce": return AlgorithmDeviceReduce(fallback_entries) elif algorithm_name == "device_scan": return AlgorithmDeviceScan(fallback_entries) elif algorithm_name == "device_scan_by_key": return AlgorithmDeviceScanByKey(fallback_entries) elif algorithm_name == "device_binary_search": return AlgorithmDeviceBinarySearch(fallback_entries) elif algorithm_name == "device_upper_bound": return AlgorithmDeviceUpperBound(fallback_entries) elif algorithm_name == "device_lower_bound": return AlgorithmDeviceLowerBound(fallback_entries) elif algorithm_name == "device_adjacent_difference": return AlgorithmDeviceAdjacentDifference(fallback_entries) elif algorithm_name == "device_adjacent_difference_inplace": return AlgorithmDeviceAdjacentDifferenceInplace(fallback_entries) elif algorithm_name == "device_adjacent_find": return AlgorithmDeviceAdjacentFind(fallback_entries) elif algorithm_name == "device_segmented_radix_sort": return AlgorithmDeviceSegmentedRadixSort(fallback_entries) elif algorithm_name == 'device_segmented_topk_air': return AlgorithmDeviceSegmentedTopkAir(fallback_entries) elif algorithm_name == "device_transform": return AlgorithmDeviceTransform(fallback_entries) elif algorithm_name == "device_topk_air": return AlgorithmDeviceTopkAir(fallback_entries) elif algorithm_name == "device_partition_two_way_predicate": return AlgorithmDevicePartitionTwoWayPredicate(fallback_entries) elif algorithm_name == "device_partition_two_way_flag": return AlgorithmDevicePartitionTwoWayFlag(fallback_entries) elif algorithm_name == "device_partition_flag": return AlgorithmDevicePartitionFlag(fallback_entries) elif algorithm_name == "device_partition_predicate": return AlgorithmDevicePartitionPredicate(fallback_entries) elif algorithm_name == "device_partition_three_way": return AlgorithmDevicePartitionThreeWay(fallback_entries) elif algorithm_name == "device_select_flag": return AlgorithmDeviceSelectFlag(fallback_entries) elif algorithm_name == "device_select_predicate": return AlgorithmDeviceSelectPredicate(fallback_entries) elif algorithm_name == "device_select_predicated_flag": return AlgorithmDeviceSelectPredicatedFlag(fallback_entries) elif algorithm_name == "device_select_unique": return AlgorithmDeviceSelectUnique(fallback_entries) elif algorithm_name == "device_select_unique_by_key": return AlgorithmDeviceSelectUniqueByKey(fallback_entries) elif algorithm_name == "device_reduce_by_key": return AlgorithmDeviceReduceByKey(fallback_entries) elif algorithm_name == "device_find_first_of": return AlgorithmDeviceFindFirstOf(fallback_entries) elif algorithm_name == "device_run_length_encode": return AlgorithmDeviceRunLengthEncode(fallback_entries) elif algorithm_name == "device_run_length_encode_non_trivial": return AlgorithmDeviceRunLengthEncodeNonTrivial(fallback_entries) elif algorithm_name == "device_merge": return AlgorithmDeviceMerge(fallback_entries) else: raise ( NotSupportedError(f'Algorithm "{algorithm_name}" is not supported (yet)') ) class TuningDataManager: """ Aggregates the data from multiple tuning files containing single tuning runs with different configurations. One file may contain data for multiple algorithms """ def __init__(self, fallback_config_file: str): self.algorithms: Dict[str, Algorithm] = {} abs_path_to_script_dir: str = os.path.dirname(os.path.abspath(__file__)) self.abs_path_to_template: str = os.path.join( abs_path_to_script_dir, "config_template" ) self.fallback_config_file: str = fallback_config_file self.fallback_entries: List[FallbackCase] = self.__load_fallback_entries() def __load_fallback_entries(self) -> List[FallbackCase]: """ Reads in fallback json file to list of dictionaries """ with open(self.fallback_config_file, "r") as f: raw_fallback_entries = json.load(f)["fallback_cases"] fallback_entries: List[FallbackCase] = [] for fallback_settings_entry in raw_fallback_entries: if ( "based_on_type" not in fallback_settings_entry or "sizeof_max_inclusive" not in fallback_settings_entry or "sizeof_min_exclusive" not in fallback_settings_entry or "is_floating_point" not in fallback_settings_entry ): raise ( ValueError( f'Fallback entry "{raw_fallback_entries}" does not have all required fields' ) ) fallback_entries.append( FallbackCase( ( None if "algo_regex" not in fallback_settings_entry else fallback_settings_entry["algo_regex"] ), fallback_settings_entry["based_on_type"], fallback_settings_entry["sizeof_max_inclusive"], fallback_settings_entry["sizeof_min_exclusive"], fallback_settings_entry["is_floating_point"], ) ) return fallback_entries def __get_target_architecture_from_context(self, tuning_run): """ Uses the tuning run context embedded into the tuning output json to retrieve the targeted architecture """ name_from_context = tuning_run["context"]["hdp_gcn_arch_name"].split(":")[0] if name_from_context in TARGET_ARCHITECTURES: return f"target_arch::{name_from_context}" else: raise RuntimeError(f"ERROR: unknown hdp_gcn_arch_name: {name_from_context}") def __get_single_tuning(self, single_tuning): """ Enriches the tuning the data in single_tuning with the information stored in the actual name of the particular tuning run This information contains the different settings the tuning has been executed with which will be used to create the customized configuration case. """ # google tuning may postfix the JSON name: extract the '{...}' substring tokenized_name = re.match(r"{.*}", single_tuning["name"]).group(0) tokenized_name = json.loads(tokenized_name) if not tokenized_name: raise RuntimeError( f"ERROR: cannot parse JSON from: \"{single_tuning['name']}\"" ) return dict(single_tuning, **tokenized_name) def add_run(self, cache_file_path: str): with open(cache_file_path, "r") as f: cache_data = json.load(f) try: print(f'INFO: Processing "{cache_file_path}"') arch_name = cache_data["arch_name"].split(":")[0] if arch_name not in TARGET_ARCHITECTURES: raise RuntimeError(f"ERROR: unknown architecture: {arch_name}") arch = f"target_arch::{arch_name}" algorithm_name = cache_data["algo_name"] if algorithm_name not in self.algorithms: self.algorithms[algorithm_name] = create_algorithm( algorithm_name, self.fallback_entries ) self.algorithms[algorithm_name].add_measurement(cache_data, arch) print(f'INFO: Successfully processed file "{cache_file_path}"') except Exception as error: print( f'WARNING: Could not process file "{cache_file_path}": {error}', file=sys.stderr, ) def write_configs_to_files(self, base_dir: str): if not self.algorithms: raise KeyError("No suitable files to process") for algo_name, algo in self.algorithms.items(): config = algo.create_config_file_content() path_str = os.path.join(base_dir, f"{algo_name}.hpp") with open(path_str, "w") as outfile: outfile.write(config) def main(): parser = argparse.ArgumentParser( description="Tool for generating optimized launch parameters for rocPRIM based on Kernel Tuner results" ) parser.add_argument( "-r", "--algo_regex", type=str, required=True, help="Regex pattern for algorithms to process", ) parser.add_argument( "-i", "--input_basedir", type=str, required=True, help="Input directory where tuning results are", ) parser.add_argument( "-o", "--out_basedir", type=str, required=True, help="Output directory for configuration files", ) parser.add_argument( "-c", "--fallback_configuration", type=str, default="fallback_config.json", help="Fallback configuration file path", ) args = parser.parse_args() tuning_manager = TuningDataManager(args.fallback_configuration) algo_pattern = re.compile(args.algo_regex.strip()) algo_output_dir = args.input_basedir for filename in os.listdir(algo_output_dir): if filename.endswith("_cache.json"): file_path = os.path.join(algo_output_dir, filename) try: with open(file_path, "r") as f: cache_data = json.load(f) if "algo_name" in cache_data and algo_pattern.match( cache_data["algo_name"] ): tuning_manager.add_run(file_path) except json.JSONDecodeError: print(f"Warning: Could not parse JSON file: {file_path}") except Exception as e: print(f"Error processing file {file_path}: {str(e)}") tuning_manager.write_configs_to_files(args.out_basedir) if __name__ == "__main__": main()