1
0
mirror of https://github.com/ARM-software/workload-automation.git synced 2025-03-25 03:59:11 +00:00
Jonathan Paynter dc64152188 configuration, types: Add support for sweeps
Sweeps are handled by one of three handlers, ``autofreq``, ``range`` and
``autoparam``. ``range`` allows the user to manually specify values for
a parameter in the agenda; ``autofreq`` allows cpu frequencies to be
swept through, and autoparam allows other parameters to be swept
through.

autoparam requries the parameter to have ``allowed_values`` specified,
and the sweep specification must supply the plugin name and parameter
name as:
``
	sweep(autoparam):
		param: <param name>
		plugin: <plugin name>
``

autofreq can be used with an empty value:
``
	sweep(autofreq):
``
to sweep through all values of ``frequency`` by default, although if
``plugin`` is specified, other parameters within cpufreq can be swept
through:
``
	sweep(autofreq):
		param: cpu0_frequency
``

For either of the above 'automatic' sweeps, a minimum and/or maximum can
be specified to limit the values generated in the sweep:
``
	sweep(autofreq):
		param: cpu0_frequency
		min: 1000000
		max: 1700000
``

``range`` sweeps support two specification syntaxes: one for manual
specification of a list, and another for a range of values. These sweeps
only accept the plugin name as a parameter:
``
	sweep(range):
		threads: [1, 3, 5, 7]
``
or
``
	sweep(range):
		threads: 1-8,2
		# start-stop[,step]
``
These both produce the same result. step is optional and defaults to 1.

Any sweep specified in a workload entry is equivalent to manually
specifying individual workload entries, each with one of the possible
values of the sweep.

Sweeps specified in sections will create the equivalent of a section
group, each section in that group with one of the sweep values. When the
group name is specified, the 'expanded' section entry will maintain that
group name, otherwise will generate its own.

If multiple sweeps are specified in one entry, then this will be
expanded to one entry for every combination of values of each sweep.
2020-09-10 18:07:16 +01:00

664 lines
24 KiB
Python

# Copyright 2015-2018 ARM Limited
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# pylint: disable=no-self-use
import os
import logging
import re
from abc import ABC, abstractmethod
from copy import deepcopy
from functools import reduce # pylint: disable=redefined-builtin
from math import inf
from devlib.utils.types import identifier
from wa.framework.configuration.core import JobSpec
from wa.framework.exception import ConfigError
from wa.utils import log
from wa.utils.serializer import json, read_pod, SerializerSyntaxError
from wa.utils.types import toggle_set, counter, sweep
from wa.utils.misc import merge_config_values, isiterable
logger = logging.getLogger('config')
class ConfigParser(object):
def load_from_path(self, state, filepath):
raw, includes = _load_file(filepath, "Config")
self.load(state, raw, filepath)
return includes
def load(self, state, raw, source, wrap_exceptions=True): # pylint: disable=too-many-branches
logger.debug('Parsing config from "{}"'.format(source))
log.indent()
try:
state.plugin_cache.add_source(source)
if 'run_name' in raw:
msg = '"run_name" can only be specified in the config '\
'section of an agenda'
raise ConfigError(msg)
if 'id' in raw:
raise ConfigError('"id" cannot be set globally')
merge_augmentations(raw)
# Get WA core configuration
for cfg_point in state.settings.configuration.values():
value = pop_aliased_param(cfg_point, raw)
if value is not None:
logger.debug('Setting meta "{}" to "{}"'.format(cfg_point.name, value))
state.settings.set(cfg_point.name, value)
# Get run specific configuration
for cfg_point in state.run_config.configuration.values():
value = pop_aliased_param(cfg_point, raw)
if value is not None:
logger.debug('Setting run "{}" to "{}"'.format(cfg_point.name, value))
state.run_config.set(cfg_point.name, value)
# Get global job spec configuration
for cfg_point in JobSpec.configuration.values():
value = pop_aliased_param(cfg_point, raw)
if value is not None:
logger.debug('Setting global "{}" to "{}"'.format(cfg_point.name, value))
state.jobs_config.set_global_value(cfg_point.name, value)
for name, values in raw.items():
# Assume that all leftover config is for a plug-in or a global
# alias it is up to PluginCache to assert this assumption
logger.debug('Caching "{}" with "{}"'.format(identifier(name), values))
state.plugin_cache.add_configs(identifier(name), values, source)
except ConfigError as e:
if wrap_exceptions:
raise ConfigError('Error in "{}":\n{}'.format(source, str(e)))
else:
raise e
finally:
log.dedent()
class AgendaParser(object):
def load_from_path(self, state, filepath):
raw, includes = _load_file(filepath, 'Agenda')
self.load(state, raw, filepath)
return includes
def load(self, state, raw, source):
logger.debug('Parsing agenda from "{}"'.format(source))
log.indent()
try:
if not isinstance(raw, dict):
raise ConfigError('Invalid agenda, top level entry must be a dict')
sections = self._pop_sections(raw)
global_workloads = self._pop_workloads(raw)
if not global_workloads:
msg = 'No jobs avaliable. Please ensure you have specified at '\
'least one workload to run.'
raise ConfigError(msg)
sect_ids, wkl_ids = self._collect_ids(sections, global_workloads)
self._populate_and_validate_config(state, raw, source, sect_ids)
if raw:
msg = 'Invalid top level agenda entry(ies): "{}"'
raise ConfigError(msg.format('", "'.join(list(raw.keys()))))
self._process_global_workloads(state, global_workloads, wkl_ids)
self._process_sections(state, sections, sect_ids, wkl_ids)
state.agenda = source
except (ConfigError, SerializerSyntaxError) as e:
raise ConfigError('Error in "{}":\n\t{}'.format(source, str(e)))
finally:
log.dedent()
def _populate_and_validate_config(self, state, raw, source, sect_ids):
for name in ['config', 'global']:
entry = raw.pop(name, None)
if entry is None:
continue
if not isinstance(entry, dict):
msg = 'Invalid entry "{}" - must be a dict'
raise ConfigError(msg.format(name))
# Want to take this entry and add any sweeps in a section
self._extract_global_sweeps(state, entry, sect_ids)
if 'run_name' in entry:
value = entry.pop('run_name')
logger.debug('Setting run name to "{}"'.format(value))
state.run_config.set('run_name', value)
state.load_config(entry, '{}/{}'.format(source, name))
def _pop_sections(self, raw):
sections = raw.pop("sections", [])
if not isinstance(sections, list):
raise ConfigError('Invalid entry "sections" - must be a list')
for section in sections:
if not hasattr(section, 'items'):
raise ConfigError('Invalid section "{}" - must be a dict'.format(section))
return sections
def _pop_workloads(self, raw):
workloads = raw.pop("workloads", [])
if not isinstance(workloads, list):
raise ConfigError('Invalid entry "workloads" - must be a list')
return workloads
def _collect_ids(self, sections, global_workloads):
seen_section_ids = set()
seen_workload_ids = set()
for workload in global_workloads:
workload = _get_workload_entry(workload)
_collect_valid_id(workload.get("id"), seen_workload_ids, "workload")
for section in sections:
_collect_valid_id(section.get("id"), seen_section_ids, "section")
for workload in section["workloads"] if "workloads" in section else []:
workload = _get_workload_entry(workload)
_collect_valid_id(workload.get("id"), seen_workload_ids,
"workload")
return seen_section_ids, seen_workload_ids
def _process_global_workloads(self, state, global_workloads, seen_wkl_ids):
for workload_entry in global_workloads:
find_sweeps(workload_entry, convert=True) # Replace sweep text with sweep type
workload = _process_workload_entry(workload_entry, seen_wkl_ids,
state.jobs_config)
state.jobs_config.add_workload(workload)
def _process_sections(self, state, sections, seen_sect_ids, seen_wkl_ids):
for section in sections:
find_sweeps(section, convert=True) # Replace sweep text with sweep type
workloads = []
for workload_entry in section.pop("workloads", []):
workload = _process_workload_entry(workload_entry, seen_wkl_ids,
state.jobs_config)
workloads.append(workload)
if 'params' in section:
if 'runtime_params' in section:
msg = 'both "params" and "runtime_params" specified in a '\
'section: "{}"'
raise ConfigError(msg.format(json.dumps(section, indent=None)))
section['runtime_params'] = section.pop('params')
group = section.pop('group', None)
section = _construct_valid_entry(section, seen_sect_ids,
"s", state.jobs_config)
state.jobs_config.add_section(section, workloads, group)
def _extract_global_sweeps(self, state, entry, seen_sect_ids):
# Find any sweeps in global config
# If there are any sweeps, add them as config inside a new section
sweep_keychains = find_sweeps(entry, convert=True)
extracted_config = {}
for keychain in sweep_keychains:
# For each sweep found
structure = entry
for key in keychain[:-1]:
structure = structure[key]
# Pop the sweep
sweep = structure.pop(keychain[-1])
# Add the sweep location to the new config
dictn = extracted_config
for key in keychain[:-1]:
dictn = dictn.setdefault(key, {})
dictn[keychain[-1]] = sweep
if extracted_config:
# Ensure unique name
id_ctr = counter('global_sweeps')
extracted_config['id'] = 'global_sweeps{}'.format(
'' if id_ctr == 1 else '_{}'.format(id_ctr)
)
grp_ctr = counter('global_sweeps')
group = 'global_sweeps{}'.format(
'' if grp_ctr == 1 else '_{}'.format(grp_ctr)
)
glbl_sweeps = _construct_valid_entry(
extracted_config,
seen_sect_ids,
None,
state.jobs_config
)
state.jobs_config.add_section(glbl_sweeps, [], group)
########################
### Helper functions ###
########################
def pop_aliased_param(cfg_point, d, default=None):
"""
Given a ConfigurationPoint and a dict, this function will search the dict for
the ConfigurationPoint's name/aliases. If more than one is found it will raise
a ConfigError. If one (and only one) is found then it will return the value
for the ConfigurationPoint. If the name or aliases are present in the dict it will
return the "default" parameter of this function.
"""
aliases = [cfg_point.name] + cfg_point.aliases
alias_map = [a for a in aliases if a in d]
if len(alias_map) > 1:
raise ConfigError('Duplicate entry: {}'.format(aliases))
elif alias_map:
return d.pop(alias_map[0])
else:
return default
def _load_file(filepath, error_name):
if not os.path.isfile(filepath):
raise ValueError("{} does not exist".format(filepath))
try:
raw = read_pod(filepath)
includes = _process_includes(raw, filepath, error_name)
except SerializerSyntaxError as e:
raise ConfigError('Error parsing {} {}: {}'.format(error_name, filepath, e))
if not isinstance(raw, dict):
message = '{} does not contain a valid {} structure; top level must be a dict.'
raise ConfigError(message.format(filepath, error_name))
return raw, includes
def _process_includes(raw, filepath, error_name):
if not raw:
return []
source_dir = os.path.dirname(filepath)
included_files = []
replace_value = None
if hasattr(raw, 'items'):
for key, value in raw.items():
if key == 'include#':
include_path = os.path.expanduser(os.path.join(source_dir, value))
included_files.append(include_path)
replace_value, includes = _load_file(include_path, error_name)
included_files.extend(includes)
elif hasattr(value, 'items') or isiterable(value):
includes = _process_includes(value, filepath, error_name)
included_files.extend(includes)
elif isiterable(raw):
for element in raw:
if hasattr(element, 'items') or isiterable(element):
includes = _process_includes(element, filepath, error_name)
included_files.extend(includes)
if replace_value is not None:
del raw['include#']
for key, value in replace_value.items():
raw[key] = merge_config_values(value, raw.get(key, None))
return included_files
def merge_augmentations(raw):
"""
Since, from configuration perspective, output processors and instruments are
handled identically, the configuration entries are now interchangeable. E.g. it is
now valid to specify a output processor in an instruments list. This is to make things
easier for the users, as, from their perspective, the distinction is somewhat arbitrary.
For backwards compatibility, both entries are still valid, and this
function merges them together into a single "augmentations" set, ensuring
that there are no conflicts between the entries.
"""
cfg_point = JobSpec.configuration['augmentations']
names = [cfg_point.name, ] + cfg_point.aliases
entries = []
for n in names:
if n not in raw:
continue
value = raw.pop(n)
try:
entries.append(toggle_set(value))
except TypeError as exc:
msg = 'Invalid value "{}" for "{}": {}'
raise ConfigError(msg.format(value, n, exc))
# Make sure none of the specified aliases conflict with each other
to_check = [e for e in entries]
while len(to_check) > 1:
check_entry = to_check.pop()
for e in to_check:
conflicts = check_entry.conflicts_with(e)
if conflicts:
msg = '"{}" and "{}" have conflicting entries: {}'
conflict_string = ', '.join('"{}"'.format(c.strip("~"))
for c in conflicts)
raise ConfigError(msg.format(check_entry, e, conflict_string))
if entries:
raw['augmentations'] = reduce(lambda x, y: x.union(y), entries)
def _pop_aliased(d, names, entry_id):
name_count = sum(1 for n in names if n in d)
if name_count > 1:
names_list = ', '.join(names)
msg = 'Invalid workload entry "{}": at most one of ({}}) must be specified.'
raise ConfigError(msg.format(entry_id, names_list))
for name in names:
if name in d:
return d.pop(name)
return None
def _construct_valid_entry(raw, seen_ids, prefix, jobs_config):
workload_entry = {}
# Generate an automatic ID if the entry doesn't already have one
if 'id' not in raw:
while True:
new_id = '{}{}'.format(prefix, counter(name=prefix))
if new_id not in seen_ids:
break
workload_entry['id'] = new_id
seen_ids.add(new_id)
else:
workload_entry['id'] = raw.pop('id')
# Process instruments
merge_augmentations(raw)
# Validate all workload_entry
for name, cfg_point in JobSpec.configuration.items():
value = pop_aliased_param(cfg_point, raw)
if value is not None:
if isinstance(value, sweep) and not value.auto:
# If values have already been specified, check the kind
value = sweep(values=map(cfg_point.kind, value))
else:
value = cfg_point.kind(value)
cfg_point.validate_value(name, value)
workload_entry[name] = value
if "augmentations" in workload_entry:
if '~~' in workload_entry['augmentations']:
msg = '"~~" can only be specfied in top-level config, and not for individual workloads/sections'
raise ConfigError(msg)
jobs_config.update_augmentations(workload_entry['augmentations'])
# error if there are unknown workload_entry
if raw:
msg = 'Invalid entry(ies) in "{}": "{}"'
raise ConfigError(msg.format(workload_entry['id'], ', '.join(list(raw.keys()))))
return workload_entry
def _collect_valid_id(entry_id, seen_ids, entry_type):
if entry_id is None:
return
entry_id = str(entry_id)
if entry_id in seen_ids:
raise ConfigError('Duplicate {} ID "{}".'.format(entry_type, entry_id))
# "-" is reserved for joining section and workload IDs
if "-" in entry_id:
msg = 'Invalid {} ID "{}"; IDs cannot contain a "-"'
raise ConfigError(msg.format(entry_type, entry_id))
if entry_id == "global":
msg = 'Invalid {} ID "global"; is a reserved ID'
raise ConfigError(msg.format(entry_type))
seen_ids.add(entry_id)
def _get_workload_entry(workload):
if isinstance(workload, str):
workload = {'name': workload}
elif not isinstance(workload, dict):
raise ConfigError('Invalid workload entry: "{}"')
return workload
def _process_workload_entry(workload, seen_workload_ids, jobs_config):
workload = _get_workload_entry(workload)
workload = _construct_valid_entry(workload, seen_workload_ids,
"wk", jobs_config)
if "workload_name" not in workload:
raise ConfigError('No workload name specified in entry {}'.format(workload['id']))
return workload
def find_sweeps(raw, convert=False):
keychain = []
if isinstance(raw, dict):
to_convert = {}
for k, v in raw.items():
if _is_sweep(k):
if convert:
sweep = _create_sweep(_sweep_handler_name(k), raw.get(k))
to_convert[k] = sweep
keychain.append((sweep.param_name, ))
else:
keychain.append((k,))
elif isinstance(v, dict) or isinstance(v, list):
subkeys = find_sweeps(v, convert=convert)
keychain.extend(map(lambda subkey: (k,) + subkey, subkeys))
for k, sweep in to_convert.items():
raw[sweep.param_name] = sweep
raw.pop(k)
elif isinstance(raw, list):
for index, v in enumerate(raw):
if not isinstance(v, dict):
continue
subkeys = find_sweeps(v, convert=convert)
keychain.extend(map(lambda subkey: (index,) + subkey, subkeys))
return keychain
def _is_sweep(name):
return name[:6] == 'sweep(' and name[-1] == ')'
def _sweep_handler_name(name):
return name[6:-1]
range_syntax = r'([0-9]+-[0-9]+)(,[0-9]+)?'
def _create_sweep(handler_name, raw_definition):
try:
handler_kind = _sweep_handlers[handler_name]
except KeyError:
msg = '{} is not a valid sweep handler'
raise ConfigError(msg.format(handler_name))
else:
return sweep(handler=handler_kind(raw_definition))
class SweepHandler(ABC):
'''
Handles any functionality required for a sweep
'''
auto = False
def __init__(self, raw_value: dict):
self.raw = deepcopy(raw_value)
self.plugin = None
self.param_name = None
self.values = None
self.parse()
@abstractmethod
def parse(self):
"""
Extract all information required from ``self.raw``, the
value associated with the sweep key in the config
"""
pass
class RangeHandler(SweepHandler):
def parse(self):
# Only require a param name and values
# Other arguments ignored
if not len(self.raw) == 1:
msg = 'Too many entries for range sweep'
raise ConfigError(msg)
self.param_name, vals = tuple(self.raw.items())[0]
if isinstance(vals, list):
self.values = vals
if not self.values:
msg = 'At least 1 value must be specified in list for range '\
'sweep {}'
raise ConfigError(msg.format(self.param_name))
elif isinstance(vals, str):
vals = ''.join(vals.split())
match = re.match(range_syntax, vals)
if match:
start, stop = match[1].split('-')
step = match[2][1:] if match[2] else 1
start, stop, step = int(start), int(stop), int(step)
self.values = list(range(start, stop, step))
else:
msg = 'Invalid range sweep format for param {}'
raise ConfigError(msg.format(self.param_name))
else:
msg = 'Invalid range sweep format for param {}'
raise ConfigError(msg.format(self.param_name))
class AutoSweepHandler(SweepHandler):
auto = True
def __init__(self, raw_value):
self.min = None
self.max = None
super().__init__(raw_value)
def parse(self):
# This ``parse`` method, if called, must be called *after* any child
# parse methods that pop from raw, as it assumes that raw should be
# empty by the end
if self.raw:
try:
min = self.raw.pop('min', None)
max = self.raw.pop('max', None)
if min is not None:
self.min = float(min)
if max is not None:
self.max = float(max)
except ValueError:
min_msg = 'minimum {} '.format(min) if min else ''
max_msg = 'maximum {} '.format(max) if max else ''
connective = 'and ' if min_msg and max_msg else ''
raise ConfigError('Sweep {}{}{}must be numeric'.format(
min_msg, connective, max_msg
))
else:
# If either is specified, both must be
if min and not max:
self.max = inf
elif max and not min:
self.min = -inf
self.param_name = self.raw.pop('param', None)
self.plugin = self.raw.pop('plugin', None)
if self.raw:
msg = 'Too many arguments to sweep definition'
raise ConfigError(msg)
@abstractmethod
def resolve_auto_sweep(self, tm, pc):
"""
Convert the auto sweep found in ``self.raw`` to
a list of values, to be stored in ``self.values``
:param tm: TargetManager
:param pc: PluginCache
"""
pass
class FreqHandler(AutoSweepHandler):
def parse(self):
super().parse()
self.param_name = self.param_name if self.param_name is not None else 'frequency'
def resolve_auto_sweep(self, tm, pc):
freq_cfg = tm.rpm.get_cfg_point(self.param_name)
allowed_values = None
if hasattr(freq_cfg, 'kind') and hasattr(freq_cfg.kind, 'values'):
allowed_values = freq_cfg.kind.values
elif hasattr(freq_cfg, 'allowed_values'):
allowed_values = freq_cfg.allowed_values
else:
msg = 'Runtime config parameter {} can not be swept'
raise ConfigError(msg.format(self.param_name))
if self.min:
self.values = list(filter(lambda x: self.min < x < self.max, allowed_values))
else:
self.values = allowed_values
class ParamHandler(AutoSweepHandler):
"""
For auto sweeps of any parameter that specifies
``allowed_values``
"""
def parse(self):
super().parse()
# Both plugin and param name must be specified
if self.plugin is None or self.param_name is None:
msg = 'autoparam sweeps require both the plugin ' \
'and parameter name to be specified'
raise ConfigError(msg)
def resolve_auto_sweep(self, tm, pc):
params = pc.get_plugin_parameters(self.plugin)
cfg_pt = params[self.param_name]
if cfg_pt.allowed_values is None:
msg = 'Parameter \'{}\' does not specify allowed values to sweep'
raise ConfigError(msg.format(self.param_name))
# If min is not None, then the values are numeric, and both min and max
# will be numeric values
if self.min is not None:
self.values = list(filter(lambda x: self.min < x < self.max, cfg_pt.allowed_values))
else:
self.values = cfg_pt.allowed_values
_sweep_handlers = {
'range': RangeHandler,
'autofreq': FreqHandler,
'autoparam': ParamHandler,
}