#!/usr/bin/env python3 import csv import io import json import numpy as np import os import re from scipy import optimize from sklearn.metrics import r2_score import struct import sys import tarfile from multiprocessing import Pool from automata import PTA from functions import analytic from functions import AnalyticFunction from utils import vprint, is_numeric, soft_cast_int, param_slice_eq, compute_param_statistics arg_support_enabled = True def running_mean(x: np.ndarray, N: int) -> np.ndarray: """ Compute running average. arguments: x -- NumPy array N -- how many items to average """ cumsum = np.cumsum(np.insert(x, 0, 0)) return (cumsum[N:] - cumsum[:-N]) / N def gplearn_to_function(function_str: str): """ Convert gplearn-style function string to Python function. Takes a function string like "mul(add(X0, X1), X2)" and returns a Python function implementing the specified behaviour, e.g. "lambda x, y, z: (x + y) * z". Supported functions: add -- x + y sub -- x - y mul -- x * y div -- x / y if |y| > 0.001, otherwise 1 sqrt -- sqrt(|x|) log -- log(|x|) if |x| > 0.001, otherwise 0 inv -- 1 / x if |x| > 0.001, otherwise 0 """ eval_globals = { 'add' : lambda x, y : x + y, 'sub' : lambda x, y : x - y, 'mul' : lambda x, y : x * y, 'div' : lambda x, y : np.divide(x, y) if np.abs(y) > 0.001 else 1., 'sqrt': lambda x : np.sqrt(np.abs(x)), 'log' : lambda x : np.log(np.abs(x)) if np.abs(x) > 0.001 else 0., 'inv' : lambda x : 1. / x if np.abs(x) > 0.001 else 0., } last_arg_index = 0 for i in range(0, 100): if function_str.find('X{:d}'.format(i)) >= 0: last_arg_index = i arg_list = [] for i in range(0, last_arg_index+1): arg_list.append('X{:d}'.format(i)) eval_str = 'lambda {}, *whatever: {}'.format(','.join(arg_list), function_str) print(eval_str) return eval(eval_str, eval_globals) def _arg_name(arg_index: int) -> str: return '~arg{:02}'.format(arg_index) def append_if_set(aggregate: dict, data: dict, key: str): """Append data[key] to aggregate if key in data.""" if key in data: aggregate.append(data[key]) def mean_or_none(arr): """Compute mean of NumPy array arr, return -1 if empty.""" if len(arr): return np.mean(arr) return -1 def aggregate_measures(aggregate: float, actual: list) -> dict: """ Calculate error measures for model value on data list. arguments: aggregate -- model value (float or int) actual -- real-world / reference values (list of float or int) return value: See regression_measures """ aggregate_array = np.array([aggregate] * len(actual)) return regression_measures(aggregate_array, np.array(actual)) def regression_measures(predicted: np.ndarray, actual: np.ndarray): """ Calculate error measures by comparing model values to reference values. arguments: predicted -- model values (np.ndarray) actual -- real-world / reference values (np.ndarray) Returns a dict containing the following measures: mae -- Mean Absolute Error mape -- Mean Absolute Percentage Error, if all items in actual are non-zero (NaN otherwise) smape -- Symmetric Mean Absolute Percentage Error, if no 0,0-pairs are present in actual and predicted (NaN otherwise) msd -- Mean Square Deviation rmsd -- Root Mean Square Deviation ssr -- Sum of Squared Residuals rsq -- R^2 measure, see sklearn.metrics.r2_score count -- Number of values """ if type(predicted) != np.ndarray: raise ValueError('first arg must be ndarray, is {}'.format(type(predicted))) if type(actual) != np.ndarray: raise ValueError('second arg must be ndarray, is {}'.format(type(actual))) deviations = predicted - actual #mean = np.mean(actual) if len(deviations) == 0: return {} measures = { 'mae' : np.mean(np.abs(deviations), dtype=np.float64), 'msd' : np.mean(deviations**2, dtype=np.float64), 'rmsd' : np.sqrt(np.mean(deviations**2), dtype=np.float64), 'ssr' : np.sum(deviations**2, dtype=np.float64), 'rsq' : r2_score(actual, predicted), 'count' : len(actual), } #rsq_quotient = np.sum((actual - mean)**2, dtype=np.float64) * np.sum((predicted - mean)**2, dtype=np.float64) if np.all(actual != 0): measures['mape'] = np.mean(np.abs(deviations / actual)) * 100 # bad measure else: measures['mape'] = np.nan if np.all(np.abs(predicted) + np.abs(actual) != 0): measures['smape'] = np.mean(np.abs(deviations) / (( np.abs(predicted) + np.abs(actual)) / 2 )) * 100 else: measures['smape'] = np.nan #if np.all(rsq_quotient != 0): # measures['rsq'] = (np.sum((actual - mean) * (predicted - mean), dtype=np.float64)**2) / rsq_quotient return measures class KeysightCSV: """Simple loader for Keysight CSV data, as exported by the windows software.""" def __init__(self): """Create a new KeysightCSV object.""" pass def load_data(self, filename: str): """ Load log data from filename, return timestamps and currents. Returns two one-dimensional NumPy arrays: timestamps and corresponding currents. """ with open(filename) as f: for i, _ in enumerate(f): pass timestamps = np.ndarray((i-3), dtype=float) currents = np.ndarray((i-3), dtype=float) # basically seek back to start with open(filename) as f: for _ in range(4): next(f) reader = csv.reader(f, delimiter=',') for i, row in enumerate(reader): timestamps[i] = float(row[0]) currents[i] = float(row[2]) * -1 return timestamps, currents def by_name_to_by_param(by_name: dict): """ Convert aggregation by name to aggregation by name and parameter values. """ by_param = dict() for name in by_name.keys(): for i, parameters in enumerate(by_name[name]['param']): param_key = (name, tuple(parameters)) if param_key not in by_param: by_param[param_key] = dict() for key in by_name[name].keys(): by_param[param_key][key] = list() by_param[param_key]['attributes'] = by_name[name]['attributes'] # special case for PTA models if 'isa' in by_name[name]: by_param[param_key]['isa'] = by_name[name]['isa'] for attribute in by_name[name]['attributes']: by_param[param_key][attribute].append(by_name[name][attribute][i]) return by_param def _xv_partitions_kfold(length, num_slices): pairs = [] indexes = np.arange(length) for i in range(0, num_slices): training = np.delete(indexes, slice(i, None, num_slices)) validation = indexes[i::num_slices] pairs.append((training, validation)) return pairs def _xv_partition_montecarlo(length): shuffled = np.random.permutation(np.arange(length)) border = int(length * float(2) / 3) training = shuffled[:border] validation = shuffled[border:] return (training, validation) class CrossValidator: """ Cross-Validation helper for model generation. Given a set of measurements and a model class, it will partition the data into training and validation sets, train the model on the training set, and assess its quality on the validation set. This is repeated several times depending on cross-validation algorithm and configuration. Reports the mean model error over all cross-validation runs. """ def __init__(self, model_class, by_name, parameters, arg_count): """ Create a new CrossValidator object. Does not perform cross-validation yet. arguments: model_class -- model class/type used for model synthesis, e.g. PTAModel or AnalyticModel. model_class must have a constructor accepting (by_name, parameters, arg_count, verbose = False) and provide an assess method. by_name -- measurements aggregated by state/transition/function/... name. Layout: by_name[name][attribute] = list of data. Additionally, by_name[name]['attributes'] must be set to the list of attributes, e.g. ['power'] or ['duration', 'energy']. """ self.model_class = model_class self.by_name = by_name self.names = sorted(by_name.keys()) self.parameters = sorted(parameters) self.arg_count = arg_count def montecarlo(self, model_getter, count = 200): """ Perform Monte Carlo cross-validation and return average model quality. The by_name data is randomly divided into 2/3 training and 1/3 validation. After creating a model for the training set, the model type returned by model_getter is evaluated on the validation set. This is repeated count times (defaulting to 200); the average of all measures is returned to the user. arguments: model_getter -- function with signature (model_object) -> model, e.g. lambda m: m.get_fitted()[0] to evaluate the parameter-aware model with automatic parameter detection. count -- number of validation runs to perform, defaults to 200 return value: dict of model quality measures. { 'by_name' : { for each name: { for each attribute: { 'mae' : mean of all mean absolute errors 'mae_list' : list of the individual MAE values encountered during cross-validation 'smape' : mean of all symmetric mean absolute percentage errors 'smape_list' : list of the individual SMAPE values encountered during cross-validation } } } } """ ret = { 'by_name' : dict() } for name in self.names: ret['by_name'][name] = dict() for attribute in self.by_name[name]['attributes']: ret['by_name'][name][attribute] = { 'mae_list': list(), 'smape_list': list() } for _ in range(count): res = self._single_montecarlo(model_getter) for name in self.names: for attribute in self.by_name[name]['attributes']: ret['by_name'][name][attribute]['mae_list'].append(res['by_name'][name][attribute]['mae']) ret['by_name'][name][attribute]['smape_list'].append(res['by_name'][name][attribute]['smape']) for name in self.names: for attribute in self.by_name[name]['attributes']: ret['by_name'][name][attribute]['mae'] = np.mean(ret['by_name'][name][attribute]['mae_list']) ret['by_name'][name][attribute]['smape'] = np.mean(ret['by_name'][name][attribute]['smape_list']) return ret def _single_montecarlo(self, model_getter): training = dict() validation = dict() for name in self.names: training[name] = { 'attributes' : self.by_name[name]['attributes'] } validation[name] = { 'attributes' : self.by_name[name]['attributes'] } if 'isa' in self.by_name[name]: training[name]['isa'] = self.by_name[name]['isa'] validation[name]['isa'] = self.by_name[name]['isa'] data_count = len(self.by_name[name]['param']) training_subset, validation_subset = _xv_partition_montecarlo(data_count) for attribute in self.by_name[name]['attributes']: self.by_name[name][attribute] = np.array(self.by_name[name][attribute]) training[name][attribute] = self.by_name[name][attribute][training_subset] validation[name][attribute] = self.by_name[name][attribute][validation_subset] # We can't use slice syntax for 'param', which may contain strings and other odd values training[name]['param'] = list() validation[name]['param'] = list() for idx in training_subset: training[name]['param'].append(self.by_name[name]['param'][idx]) for idx in validation_subset: validation[name]['param'].append(self.by_name[name]['param'][idx]) training_data = self.model_class(training, self.parameters, self.arg_count, verbose = False) training_model = model_getter(training_data) validation_data = self.model_class(validation, self.parameters, self.arg_count, verbose = False) return validation_data.assess(training_model) def _preprocess_measurement(measurement): setup = measurement['setup'] mim = MIMOSA(float(setup['mimosa_voltage']), int(setup['mimosa_shunt'])) charges, triggers = mim.load_data(measurement['content']) trigidx = mim.trigger_edges(triggers) triggers = [] cal_edges = mim.calibration_edges(running_mean(mim.currents_nocal(charges[0:trigidx[0]]), 10)) calfunc, caldata = mim.calibration_function(charges, cal_edges) vcalfunc = np.vectorize(calfunc, otypes=[np.float64]) processed_data = { 'fileno' : measurement['fileno'], 'info' : measurement['info'], 'triggers' : len(trigidx), 'first_trig' : trigidx[0] * 10, 'calibration' : caldata, 'trace' : mim.analyze_states(charges, trigidx, vcalfunc) } return processed_data class ParamStats: def __init__(self, by_name, by_param, parameter_names, arg_count, use_corrcoef = False, verbose = False): """ Compute standard deviation and correlation coefficient on parameterized data partitions. It is strongly recommended to vary all parameter values evenly. For instance, given two parameters, providing only the combinations (1, 1), (5, 1), (7, 1,) (10, 1), (1, 2), (1, 6) will lead to bogus results. It is better to provide (1, 1), (5, 1), (1, 2), (5, 2), ... (i.e. a cross product of all individual parameter values) arguments: by_name -- ground truth partitioned by state/transition name. by_name[state_or_trans][attribute] must be a list or 1-D numpy array. by_name[state_or_trans]['param'] must be a list of parameter values corresponding to the ground truth, e.g. [[1, 2, 3], ...] if the first ground truth element has the (lexically) first parameter set to 1, the second to 2 and the third to 3. by_param -- ground truth partitioned by state/transition name and parameters. by_name[(state_or_trans, *)][attribute] must be a list or 1-D numpy array. parameter_names -- list of parameter names, must have the same order as the parameter values in by_param (lexical sorting is recommended). arg_count -- dict providing the number of functions args ("local parameters") for each function. use_corrcoef -- use correlation coefficient instead of stddev heuristic for parameter detection """ self.stats = dict() self.use_corrcoef = use_corrcoef # Note: This is deliberately single-threaded. The overhead incurred # by multiprocessing is higher than the speed gained by parallel # computation of statistics measures. for state_or_tran in by_name.keys(): self.stats[state_or_tran] = dict() for attribute in by_name[state_or_tran]['attributes']: self.stats[state_or_tran][attribute] = compute_param_statistics(by_name, by_param, parameter_names, arg_count, state_or_tran, attribute, verbose = verbose) def _generic_param_independence_ratio(self, state_or_trans, attribute): """ Return the heuristic ratio of parameter independence for state_or_trans and attribute. This is not supported if the correlation coefficient is used. A value close to 1 means no influence, a value close to 0 means high probability of influence. """ statistics = self.stats[state_or_trans][attribute] if self.use_corrcoef: # not supported raise ValueError if statistics['std_static'] == 0: return 0 return statistics['std_param_lut'] / statistics['std_static'] def generic_param_dependence_ratio(self, state_or_trans, attribute): """ Return the heuristic ratio of parameter dependence for state_or_trans and attribute. This is not supported if the correlation coefficient is used. A value close to 0 means no influence, a value close to 1 means high probability of influence. """ return 1 - self._generic_param_independence_ratio(state_or_trans, attribute) def _param_independence_ratio(self, state_or_trans, attribute, param): """ Return the heuristic ratio of parameter independence for state_or_trans, attribute, and param. A value close to 1 means no influence, a value close to 0 means high probability of influence. """ statistics = self.stats[state_or_trans][attribute] if self.use_corrcoef: return 1 - np.abs(statistics['corr_by_param'][param]) if statistics['std_by_param'][param] == 0: if statistics['std_param_lut'] != 0: raise RuntimeError("wat") # In general, std_param_lut < std_by_param. So, if std_by_param == 0, std_param_lut == 0 follows. # This means that the variation of param does not affect the model quality -> no influence, return 1 return 1 return statistics['std_param_lut'] / statistics['std_by_param'][param] def param_dependence_ratio(self, state_or_trans, attribute, param): """ Return the heuristic ratio of parameter dependence for state_or_trans, attribute, and param. A value close to 0 means no influence, a value close to 1 means high probability of influence. """ return 1 - self._param_independence_ratio(state_or_trans, attribute, param) def _arg_independence_ratio(self, state_or_trans, attribute, arg_index): statistics = self.stats[state_or_trans][attribute] if self.use_corrcoef: return 1 - np.abs(statistics['corr_by_arg'][arg_index]) if statistics['std_by_arg'][arg_index] == 0: if statistics['std_param_lut'] != 0: raise RuntimeError("wat") # In general, std_param_lut < std_by_arg. So, if std_by_arg == 0, std_param_lut == 0 follows. # This means that the variation of arg does not affect the model quality -> no influence, return 1 return 1 return statistics['std_param_lut'] / statistics['std_by_arg'][arg_index] def arg_dependence_ratio(self, state_or_trans, attribute, arg_index): return 1 - self._arg_independence_ratio(state_or_trans, attribute, arg_index) # This heuristic is very similar to the "function is not much better than # median" checks in get_fitted. So far, doing it here as well is mostly # a performance and not an algorithm quality decision. # --df, 2018-04-18 def depends_on_param(self, state_or_trans, attribute, param): """Return whether attribute of state_or_trans depens on param.""" if self.use_corrcoef: return self.param_dependence_ratio(state_or_trans, attribute, param) > 0.1 else: return self.param_dependence_ratio(state_or_trans, attribute, param) > 0.5 # See notes on depends_on_param def depends_on_arg(self, state_or_trans, attribute, arg_index): """Return whether attribute of state_or_trans depens on arg_index.""" if self.use_corrcoef: return self.arg_dependence_ratio(state_or_trans, attribute, arg_index) > 0.1 else: return self.arg_dependence_ratio(state_or_trans, attribute, arg_index) > 0.5 class TimingData: """ Loader for timing model traces measured with on-board timers using `harness.OnboardTimerHarness`. Excpets a specific trace format and UART log output (as produced by generate-dfa-benchmark.py). Prunes states from output. (TODO) """ def __init__(self, filenames): """ Create a new TimingData object. Each filenames element corresponds to a measurement run. """ self.filenames = filenames.copy() self.traces_by_fileno = [] self.setup_by_fileno = [] self.preprocessed = False self._parameter_names = None self.version = 0 def _concatenate_analyzed_traces(self): self.traces = [] for trace_group in self.traces_by_fileno: for trace in trace_group: # TimingHarness logs states, but does not aggregate any data for them at the moment -> throw all states away transitions = list(filter(lambda x: x['isa'] == 'transition', trace['trace'])) self.traces.append({ 'id' : trace['id'], 'trace': transitions, }) for i, trace in enumerate(self.traces): trace['orig_id'] = trace['id'] trace['id'] = i for log_entry in trace['trace']: paramkeys = sorted(log_entry['parameter'].keys()) paramvalues = [soft_cast_int(log_entry['parameter'][x]) for x in paramkeys] if arg_support_enabled and 'args' in log_entry: paramvalues.extend(map(soft_cast_int, log_entry['args'])) if not 'param' in log_entry['offline_aggregates']: log_entry['offline_aggregates']['param'] = list() if 'duration' in log_entry['offline_aggregates']: for i in range(len(log_entry['offline_aggregates']['duration'])): log_entry['offline_aggregates']['param'].append(paramvalues) def _preprocess_0(self): for filename in self.filenames: with open(filename, 'r') as f: log_data = json.load(f) self.traces_by_fileno.append(log_data['traces']) self._concatenate_analyzed_traces() def get_preprocessed_data(self, verbose = True): """ Return a list of DFA traces annotated with timing, and parameter data. Suitable for the PTAModel constructor. See PTAModel(...) docstring for format details. """ self.verbose = verbose if self.preprocessed: return self.traces if self.version == 0: self._preprocess_0() self.preprocessed = True return self.traces def sanity_check_aggregate(aggregate): for key in aggregate: if not 'param' in aggregate[key]: raise RuntimeError('aggregate[{}][param] does not exist'.format(key)) if not 'attributes' in aggregate[key]: raise RuntimeError('aggregate[{}][attributes] does not exist'.format(key)) for attribute in aggregate[key]['attributes']: if not attribute in aggregate[key]: raise RuntimeError('aggregate[{}][{}] does not exist, even though it is contained in aggregate[{}][attributes]'.format(key, attribute, key)) param_len = len(aggregate[key]['param']) attr_len = len(aggregate[key][attribute]) if param_len != attr_len: raise RuntimeError('parameter mismatch: len(aggregate[{}][param]) == {} != len(aggregate[{}][{}]) == {}'.format(key, param_len, key, attribute, attr_len)) class RawData: """ Loader for hardware model traces measured with MIMOSA. Expects a specific trace format and UART log output (as produced by the dfatool benchmark generator). Loads data, prunes bogus measurements, and provides preprocessed data suitable for PTAModel. """ def __init__(self, filenames): """ Create a new RawData object. Each filename element corresponds to a measurement run. It must be a tar archive with the following contents: * `setup.json`: measurement setup. Must contain the keys `state_duration` (how long each state is active, in ms), `mimosa_voltage` (voltage applied to dut, in V), and `mimosa_shunt` (shunt value, in Ohm) * `src/apps/DriverEval/DriverLog.json`: PTA traces and parameters for this benchmark. Layout: List of traces, each trace has an 'id' (numeric, starting with 1) and 'trace' (list of states and transitions) element. Each trace has an even number of elements, starting with the first state (usually `UNINITIALIZED`) and ending with a transition. Each state/transition must have the members `.parameter` (parameter values, empty string or None if unknown), `.isa` ("state" or "transition") and `.name`. Each transition must additionally contain `.plan.level` ("user" or "epilogue"). Example: `[ {"id": 1, "trace": [ {"parameter": {...}, "isa": "state", "name": "UNINITIALIZED"}, ...] }, ... ] * At least one `*.mim` file. Each file corresponds to a single execution of the entire benchmark (i.e., all runs described in DriverLog.json) and starts with a MIMOSA Autocal calibration sequence. MIMOSA files are parsed by the `MIMOSA` class. """ self.filenames = filenames.copy() self.traces_by_fileno = [] self.setup_by_fileno = [] self.version = 0 self.preprocessed = False self._parameter_names = None def _state_is_too_short(self, online, offline, state_duration, next_transition): # We cannot control when an interrupt causes a state to be left if next_transition['plan']['level'] == 'epilogue': return False # Note: state_duration is stored as ms, not us return offline['us'] < state_duration * 500 def _state_is_too_long(self, online, offline, state_duration, prev_transition): # If the previous state was left by an interrupt, we may have some # waiting time left over. So it's okay if the current state is longer # than expected. if prev_transition['plan']['level'] == 'epilogue': return False # state_duration is stored as ms, not us return offline['us'] > state_duration * 1500 def _measurement_is_valid(self, processed_data): setup = self.setup_by_fileno[processed_data['fileno']] traces = self.traces_by_fileno[processed_data['fileno']] state_duration = setup['state_duration'] # Check trigger count sched_trigger_count = 0 for run in traces: sched_trigger_count += len(run['trace']) if sched_trigger_count != processed_data['triggers']: processed_data['error'] = 'got {got:d} trigger edges, expected {exp:d}'.format( got = processed_data['triggers'], exp = sched_trigger_count ) return False # Check state durations. Very short or long states can indicate a # missed trigger signal which wasn't detected due to duplicate # triggers elsewhere online_datapoints = [] for run_idx, run in enumerate(traces): for trace_part_idx in range(len(run['trace'])): online_datapoints.append((run_idx, trace_part_idx)) for offline_idx, online_ref in enumerate(online_datapoints): online_run_idx, online_trace_part_idx = online_ref offline_trace_part = processed_data['trace'][offline_idx] online_trace_part = traces[online_run_idx]['trace'][online_trace_part_idx] if self._parameter_names == None: self._parameter_names = sorted(online_trace_part['parameter'].keys()) if sorted(online_trace_part['parameter'].keys()) != self._parameter_names: processed_data['error'] = 'Offline #{off_idx:d} (online {on_name:s} @ {on_idx:d}/{on_sub:d}) has inconsistent paramete set: should be {param_want:s}, is {param_is:s}'.format( off_idx = offline_idx, on_idx = online_run_idx, on_sub = online_trace_part_idx, on_name = online_trace_part['name'], param_want = self._parameter_names, param_is = sorted(online_trace_part['parameter'].keys()) ) if online_trace_part['isa'] != offline_trace_part['isa']: processed_data['error'] = 'Offline #{off_idx:d} (online {on_name:s} @ {on_idx:d}/{on_sub:d}) claims to be {off_isa:s}, but should be {on_isa:s}'.format( off_idx = offline_idx, on_idx = online_run_idx, on_sub = online_trace_part_idx, on_name = online_trace_part['name'], off_isa = offline_trace_part['isa'], on_isa = online_trace_part['isa']) return False # Clipping in UNINITIALIZED (offline_idx == 0) can happen during # calibration and is handled by MIMOSA if offline_idx != 0 and offline_trace_part['clip_rate'] != 0: processed_data['error'] = 'Offline #{off_idx:d} (online {on_name:s} @ {on_idx:d}/{on_sub:d}) was clipping {clip:f}% of the time'.format( off_idx = offline_idx, on_idx = online_run_idx, on_sub = online_trace_part_idx, on_name = online_trace_part['name'], clip = offline_trace_part['clip_rate'] * 100, ) return False if online_trace_part['isa'] == 'state' and online_trace_part['name'] != 'UNINITIALIZED': online_prev_transition = traces[online_run_idx]['trace'][online_trace_part_idx-1] online_next_transition = traces[online_run_idx]['trace'][online_trace_part_idx+1] try: if self._state_is_too_short(online_trace_part, offline_trace_part, state_duration, online_next_transition): processed_data['error'] = 'Offline #{off_idx:d} (online {on_name:s} @ {on_idx:d}/{on_sub:d}) is too short (duration = {dur:d} us)'.format( off_idx = offline_idx, on_idx = online_run_idx, on_sub = online_trace_part_idx, on_name = online_trace_part['name'], dur = offline_trace_part['us']) return False if self._state_is_too_long(online_trace_part, offline_trace_part, state_duration, online_prev_transition): processed_data['error'] = 'Offline #{off_idx:d} (online {on_name:s} @ {on_idx:d}/{on_sub:d}) is too long (duration = {dur:d} us)'.format( off_idx = offline_idx, on_idx = online_run_idx, on_sub = online_trace_part_idx, on_name = online_trace_part['name'], dur = offline_trace_part['us']) return False except KeyError: pass # TODO es gibt next_transitions ohne 'plan' return True def _merge_measurement_into_online_data(self, measurement): online_datapoints = [] traces = self.traces_by_fileno[measurement['fileno']] for run_idx, run in enumerate(traces): for trace_part_idx in range(len(run['trace'])): online_datapoints.append((run_idx, trace_part_idx)) for offline_idx, online_ref in enumerate(online_datapoints): online_run_idx, online_trace_part_idx = online_ref offline_trace_part = measurement['trace'][offline_idx] online_trace_part = traces[online_run_idx]['trace'][online_trace_part_idx] if not 'offline' in online_trace_part: online_trace_part['offline'] = [offline_trace_part] else: online_trace_part['offline'].append(offline_trace_part) paramkeys = sorted(online_trace_part['parameter'].keys()) paramvalue = [soft_cast_int(online_trace_part['parameter'][x]) for x in paramkeys] # NB: Unscheduled transitions do not have an 'args' field set. # However, they should only be caused by interrupts, and # interrupts don't have args anyways. if arg_support_enabled and 'args' in online_trace_part: paramvalue.extend(map(soft_cast_int, online_trace_part['args'])) if not 'offline_aggregates' in online_trace_part: online_trace_part['offline_attributes'] = ['power', 'duration', 'energy'] online_trace_part['offline_aggregates'] = { 'power' : [], 'duration' : [], 'power_std' : [], 'energy' : [], 'paramkeys' : [], 'param': [], } if online_trace_part['isa'] == 'transition': online_trace_part['offline_attributes'].extend(['rel_energy_prev', 'rel_energy_next', 'timeout']) online_trace_part['offline_aggregates']['rel_energy_prev'] = [] online_trace_part['offline_aggregates']['rel_energy_next'] = [] online_trace_part['offline_aggregates']['timeout'] = [] # Note: All state/transitions are 20us "too long" due to injected # active wait states. These are needed to work around MIMOSA's # relatively low sample rate of 100 kHz (10us) and removed here. online_trace_part['offline_aggregates']['power'].append( offline_trace_part['uW_mean']) online_trace_part['offline_aggregates']['duration'].append( offline_trace_part['us'] - 20) online_trace_part['offline_aggregates']['power_std'].append( offline_trace_part['uW_std']) online_trace_part['offline_aggregates']['energy'].append( offline_trace_part['uW_mean'] * (offline_trace_part['us'] - 20)) online_trace_part['offline_aggregates']['paramkeys'].append(paramkeys) online_trace_part['offline_aggregates']['param'].append(paramvalue) if online_trace_part['isa'] == 'transition': online_trace_part['offline_aggregates']['rel_energy_prev'].append( offline_trace_part['uW_mean_delta_prev'] * (offline_trace_part['us'] - 20)) online_trace_part['offline_aggregates']['rel_energy_next'].append( offline_trace_part['uW_mean_delta_next'] * (offline_trace_part['us'] - 20)) online_trace_part['offline_aggregates']['timeout'].append( offline_trace_part['timeout']) def _concatenate_analyzed_traces(self): self.traces = [] for trace in self.traces_by_fileno: self.traces.extend(trace) for i, trace in enumerate(self.traces): trace['orig_id'] = trace['id'] trace['id'] = i def get_preprocessed_data(self, verbose = True): """ Return a list of DFA traces annotated with energy, timing, and parameter data. Each DFA trace contains the following elements: * `id`: Numeric ID, starting with 1 * `total_energy`: Total amount of energy (as measured by MIMOSA) in the entire trace * `orig_id`: Original trace ID. May differ when concatenating multiple (different) benchmarks into one analysis, i.e., when calling RawData() with more than one file argument. * `trace`: List of the individual states and transitions in this trace. Always contains an even number of elements, staring with the first state (typically "UNINITIALIZED") and ending with a transition. Each trace element (that is, an entry of the `trace` list mentioned above) contains the following elements: * `isa`: "state" or "transition" * `name`: name * `offline`: List of offline measumerents for this state/transition. Each entry contains a result for this state/transition during one benchmark execution. Entry contents: - `clip_rate`: rate of clipped energy measurements, 0 .. 1 - `raw_mean`: mean raw MIMOSA value - `raw_std`: standard deviation of raw MIMOSA value - `uW_mean`: mean power draw, uW - `uw_std`: standard deviation of power draw, uW - `us`: state/transition duration, us - `uW_mean_delta_prev`: (only for transitions) difference between uW_mean of this transition and uW_mean of previous state - `uW_mean_elta_next`: (only for transitions) difference between uW_mean of this transition and uW_mean of next state - `timeout`: (only for transitions) duration of previous state, us * `offline_aggregates`: Aggregate of `offline` entries. dict of lists, each list entry has the same length - `duration`: state/transition durations ("us"), us - `energy`: state/transition energy ("us * uW_mean"), us - `power`: mean power draw ("uW_mean"), uW - `power_std`: standard deviations of power draw ("uW_std"), uW^2 - `paramkeys`: List of lists, each sub-list contains the parameter names corresponding to the `param` entries - `param`: List of lists, each sub-list contains the parameter values for this measurement. Typically, all sub-lists are the same. - `rel_energy_prev`: (only for transitions) transition energy relative to previous state mean power, pJ - `rel_energy_next`: (only for transitions) transition energy relative to next state mean power, pJ - `timeout`: (only for transitions) duration of previous state, us * `offline_attributes`: List containing the keys of `offline_aggregates` which are meant to be part of themodel. This list ultimately decides which hardware/software attributes the model describes. If isa == state, it contains power, duration, energy If isa == transition, it contains power, duration, energy, rel_energy_prev, rel_energy_next, timeout * `online`: List of online estimations for this state/transition. Each entry contains a result for this state/transition during one benchmark execution. Entry contents for isa == state: - `time`: state/transition Entry contents for isa == transition: - `timeout`: Duration of previous state, measured using on-board timers * `parameter`: dictionary describing parameter values for this state/transition. Parameter values refer to the begin of the state/transition and do not account for changes made by the transition. * `plan`: Dictionary describing expected behaviour according to schedule / offline model. Contents for isa == state: `energy`, `power`, `time` Contents for isa == transition: `energy`, `timeout`, `level`. If level is "user", the transition is part of the regular driver API. If level is "epilogue", it is an interrupt service routine and not called explicitly. Each transition also contains: * `args`: List of arguments the corresponding function call was called with. args entries are strings which are not necessarily numeric * `code`: List of function name (first entry) and arguments (remaining entries) of the corresponding function call """ self.verbose = verbose if self.preprocessed: return self.traces if self.version == 0: self._preprocess_0() self.preprocessed = True return self.traces def _preprocess_0(self): """Load raw MIMOSA data and turn it into measurements which are ready to be analyzed.""" mim_files = [] for i, filename in enumerate(self.filenames): with tarfile.open(filename) as tf: self.setup_by_fileno.append(json.load(tf.extractfile('setup.json'))) self.traces_by_fileno.append(json.load(tf.extractfile('src/apps/DriverEval/DriverLog.json'))) for member in tf.getmembers(): _, extension = os.path.splitext(member.name) if extension == '.mim': mim_files.append({ 'content' : tf.extractfile(member).read(), 'fileno' : i, 'info' : member, 'setup' : self.setup_by_fileno[i], 'traces' : self.traces_by_fileno[i], }) with Pool() as pool: measurements = pool.map(_preprocess_measurement, mim_files) num_valid = 0 for measurement in measurements: if self._measurement_is_valid(measurement): self._merge_measurement_into_online_data(measurement) num_valid += 1 else: vprint(self.verbose, '[W] Skipping {ar:s}/{m:s}: {e:s}'.format( ar = self.filenames[measurement['fileno']], m = measurement['info'].name, e = measurement['error'])) vprint(self.verbose, '[I] {num_valid:d}/{num_total:d} measurements are valid'.format( num_valid = num_valid, num_total = len(measurements))) self._concatenate_analyzed_traces() self.preprocessing_stats = { 'num_runs' : len(measurements), 'num_valid' : num_valid } class ParallelParamFit: """ Fit a set of functions on parameterized measurements. One parameter is variale, all others are fixed. Reports the best-fitting function type for each parameter. """ def __init__(self, by_param): """Create a new ParallelParamFit object.""" self.fit_queue = [] self.by_param = by_param def enqueue(self, state_or_tran, attribute, param_index, param_name, safe_functions_enabled = False): """ Add state_or_tran/attribute/param_name to fit queue. This causes fit() to compute the best-fitting function for this model part. """ self.fit_queue.append({ 'key' : [state_or_tran, attribute, param_name], 'args' : [self.by_param, state_or_tran, attribute, param_index, safe_functions_enabled] }) def fit(self): """ Fit functions on previously enqueue data. Fitting is one in parallel with one process per core. Results can be accessed using the public ParallelParamFit.results object. """ with Pool() as pool: self.results = pool.map(_try_fits_parallel, self.fit_queue) def _try_fits_parallel(arg): return { 'key' : arg['key'], 'result' : _try_fits(*arg['args']) } def _try_fits(by_param, state_or_tran, model_attribute, param_index, safe_functions_enabled = False): functions = analytic.functions(safe_functions_enabled = safe_functions_enabled) for param_key in filter(lambda x: x[0] == state_or_tran, by_param.keys()): # We might remove elements from 'functions' while iterating over # its keys. A generator will not allow this, so we need to # convert to a list. function_names = list(functions.keys()) for function_name in function_names: function_object = functions[function_name] if is_numeric(param_key[1][param_index]) and not function_object.is_valid(param_key[1][param_index]): functions.pop(function_name, None) raw_results = {} ref_results = { 'mean' : [], 'median' : [] } results = {} for param_key in filter(lambda x: x[0] == state_or_tran, by_param.keys()): X = [] Y = [] num_valid = 0 num_total = 0 for k, v in by_param.items(): if param_slice_eq(k, param_key, param_index): num_total += 1 if is_numeric(k[1][param_index]): num_valid += 1 X.extend([float(k[1][param_index])] * len(v[model_attribute])) Y.extend(v[model_attribute]) if num_valid > 2: X = np.array(X) Y = np.array(Y) for function_name, param_function in functions.items(): raw_results[function_name] = {} error_function = param_function.error_function res = optimize.least_squares(error_function, [0, 1], args=(X, Y), xtol=2e-15) measures = regression_measures(param_function.eval(res.x, X), Y) for measure, error_rate in measures.items(): if not measure in raw_results[function_name]: raw_results[function_name][measure] = [] raw_results[function_name][measure].append(error_rate) #print(function_name, res, measures) mean_measures = aggregate_measures(np.mean(Y), Y) ref_results['mean'].append(mean_measures['rmsd']) median_measures = aggregate_measures(np.median(Y), Y) ref_results['median'].append(median_measures['rmsd']) if not len(ref_results['mean']): # Insufficient data for fitting return { 'best' : None, 'best_rmsd' : np.inf, 'results' : results, } best_fit_val = np.inf best_fit_name = None for function_name, result in raw_results.items(): if len(result) > 0: results[function_name] = {} for measure in result.keys(): results[function_name][measure] = np.mean(result[measure]) rmsd = results[function_name]['rmsd'] if rmsd < best_fit_val: best_fit_val = rmsd best_fit_name = function_name return { 'best' : best_fit_name, 'best_rmsd' : best_fit_val, 'mean_rmsd' : np.mean(ref_results['mean']), 'median_rmsd' : np.mean(ref_results['median']), 'results' : results } def _num_args_from_by_name(by_name): num_args = dict() for key, value in by_name.items(): if 'args' in value: num_args[key] = len(value['args'][0]) return num_args class AnalyticModel: u""" Parameter-aware analytic energy/data size/... model. Supports both static and parameter-based model attributes, and automatic detection of parameter-dependence. These provide measurements aggregated by (function/state/...) name and (for by_param) parameter values. Layout: dictionary with one key per name ('send', 'TX', ...) or one key per name and parameter combination (('send', (1, 2)), ('send', (2, 3)), ('TX', (1, 2)), ('TX', (2, 3)), ...). Parameter values must be ordered corresponding to the lexically sorted parameter names. Each element is in turn a dict with the following elements: - param: list of parameter values in each measurement (-> list of lists) - attributes: list of keys that should be analyzed, e.g. ['power', 'duration'] - for each attribute mentioned in 'attributes': A list with measurements. All list except for 'attributes' must have the same length. For example: parameters = ['foo_count', 'irrelevant'] by_name = { 'foo' : [1, 1, 2], 'bar' : [5, 6, 7], 'attributes' : ['foo', 'bar'], 'param' : [[1, 0], [1, 0], [2, 0]] } methods: get_static -- return static (parameter-unaware) model. get_param_lut -- return parameter-aware look-up-table model. Cannot model parameter combinations not present in by_param. get_fitted -- return parameter-aware model using fitted functions for behaviour prediction. variables: names -- function/state/... names (i.e., the keys of by_name) parameters -- parameter names stats -- ParamStats object providing parameter-dependency statistics for each name and attribute assess -- calculate model quality """ def __init__(self, by_name, parameters, arg_count = None, verbose = True): """ Create a new AnalyticModel and compute parameter statistics. parameters: `by_name`: measurements aggregated by (function/state/...) name. Layout: dictionary with one key per name ('send', 'TX', ...) or one key per name and parameter combination (('send', (1, 2)), ('send', (2, 3)), ('TX', (1, 2)), ('TX', (2, 3)), ...). Parameter values must be ordered corresponding to the lexically sorted parameter names. Each element is in turn a dict with the following elements: - param: list of parameter values in each measurement (-> list of lists) - attributes: list of keys that should be analyzed, e.g. ['power', 'duration'] - for each attribute mentioned in 'attributes': A list with measurements. All list except for 'attributes' must have the same length. For example: parameters = ['foo_count', 'irrelevant'] by_name = { 'foo' : [1, 1, 2], 'duration' : [5, 6, 7], 'attributes' : ['foo', 'duration'], 'param' : [[1, 0], [1, 0], [2, 0]] # foo_count-^ ^-irrelevant } `parameters`: List of parameter names `verbose`: Print debug/info output while generating the model? """ self.cache = dict() self.by_name = by_name self.by_param = by_name_to_by_param(by_name) self.names = sorted(by_name.keys()) self.parameters = sorted(parameters) self.verbose = verbose self._num_args = arg_count if self._num_args is None: self._num_args = _num_args_from_by_name(by_name) self.stats = ParamStats(self.by_name, self.by_param, self.parameters, self._num_args, verbose = verbose) def _get_model_from_dict(self, model_dict, model_function): model = {} for name, elem in model_dict.items(): model[name] = {} for key in elem['attributes']: try: model[name][key] = model_function(elem[key]) except RuntimeWarning: vprint(self.verbose, '[W] Got no data for {} {}'.format(name, key)) except FloatingPointError as fpe: vprint(self.verbose, '[W] Got no data for {} {}: {}'.format(name, key, fpe)) return model def param_index(self, param_name): if param_name in self.parameters: return self.parameters.index(param_name) return len(self.parameters) + int(param_name) def param_name(self, param_index): if param_index < len(self.parameters): return self.parameters[param_index] return str(param_index) def get_static(self): """ Get static model function: name, attribute -> model value. Uses the median of by_name for modeling. """ static_model = self._get_model_from_dict(self.by_name, np.median) def static_median_getter(name, key, **kwargs): return static_model[name][key] return static_median_getter def get_static_using_mean(self): """ Get static model function: name, attribute -> model value. Uses the mean of by_name for modeling. """ static_model = self._get_model_from_dict(self.by_name, np.mean) def static_mean_getter(name, key, **kwargs): return static_model[name][key] return static_mean_getter def get_param_lut(self, fallback = False): """ Get parameter-look-up-table model function: name, attribute, parameter values -> model value. The function can only give model values for parameter combinations present in by_param. By default, it raises KeyError for other values. arguments: fallback -- Fall back to the (non-parameter-aware) static model when encountering unknown parameter values """ static_model = self._get_model_from_dict(self.by_name, np.median) lut_model = self._get_model_from_dict(self.by_param, np.median) def lut_median_getter(name, key, param, arg = [], **kwargs): param.extend(map(soft_cast_int, arg)) try: return lut_model[(name, tuple(param))][key] except KeyError: if fallback: return static_model[name][key] raise return lut_median_getter def get_fitted(self, safe_functions_enabled = False): """ Get paramete-aware model function and model information function. Returns two functions: model_function(name, attribute, param=parameter values) -> model value. model_info(name, attribute) -> {'fit_result' : ..., 'function' : ... } or None """ if 'fitted_model_getter' in self.cache and 'fitted_info_getter' in self.cache: return self.cache['fitted_model_getter'], self.cache['fitted_info_getter'] static_model = self._get_model_from_dict(self.by_name, np.median) param_model = dict([[name, {}] for name in self.by_name.keys()]) paramfit = ParallelParamFit(self.by_param) for name in self.by_name.keys(): for attribute in self.by_name[name]['attributes']: for param_index, param in enumerate(self.parameters): if self.stats.depends_on_param(name, attribute, param): paramfit.enqueue(name, attribute, param_index, param, False) if arg_support_enabled and name in self._num_args: for arg_index in range(self._num_args[name]): if self.stats.depends_on_arg(name, attribute, arg_index): paramfit.enqueue(name, attribute, len(self.parameters) + arg_index, arg_index, False) paramfit.fit() for name in self.by_name.keys(): num_args = 0 if name in self._num_args: num_args = self._num_args[name] for attribute in self.by_name[name]['attributes']: fit_result = {} for result in paramfit.results: if result['key'][0] == name and result['key'][1] == attribute and result['result']['best'] != None: this_result = result['result'] if this_result['best_rmsd'] >= min(this_result['mean_rmsd'], this_result['median_rmsd']): vprint(self.verbose, '[I] Not modeling {} {} as function of {}: best ({:.0f}) is worse than ref ({:.0f}, {:.0f})'.format( name, attribute, result['key'][2], this_result['best_rmsd'], this_result['mean_rmsd'], this_result['median_rmsd'])) # See notes on depends_on_param elif this_result['best_rmsd'] >= 0.8 * min(this_result['mean_rmsd'], this_result['median_rmsd']): vprint(self.verbose, '[I] Not modeling {} {} as function of {}: best ({:.0f}) is not much better than ({:.0f}, {:.0f})'.format( name, attribute, result['key'][2], this_result['best_rmsd'], this_result['mean_rmsd'], this_result['median_rmsd'])) else: fit_result[result['key'][2]] = this_result if len(fit_result.keys()): x = analytic.function_powerset(fit_result, self.parameters, num_args) x.fit(self.by_param, name, attribute) if x.fit_success: param_model[name][attribute] = { 'fit_result': fit_result, 'function' : x } def model_getter(name, key, **kwargs): if 'arg' in kwargs and 'param' in kwargs: kwargs['param'].extend(map(soft_cast_int, kwargs['arg'])) if key in param_model[name]: param_list = kwargs['param'] param_function = param_model[name][key]['function'] if param_function.is_predictable(param_list): return param_function.eval(param_list) return static_model[name][key] def info_getter(name, key): if key in param_model[name]: return param_model[name][key] return None self.cache['fitted_model_getter'] = model_getter self.cache['fitted_info_getter'] = info_getter return model_getter, info_getter def assess(self, model_function): """ Calculate MAE, SMAPE, etc. of model_function for each by_name entry. state/transition/... name and parameter values are fed into model_function. The by_name entries of this AnalyticModel are used as ground truth and compared with the values predicted by model_function. For proper model assessments, the data used to generate model_function and the data fed into this AnalyticModel instance must be mutually exclusive (e.g. by performing cross validation). Otherwise, overfitting cannot be detected. """ detailed_results = {} for name, elem in sorted(self.by_name.items()): detailed_results[name] = {} for attribute in elem['attributes']: predicted_data = np.array(list(map(lambda i: model_function(name, attribute, param=elem['param'][i]), range(len(elem[attribute]))))) measures = regression_measures(predicted_data, elem[attribute]) detailed_results[name][attribute] = measures return { 'by_name' : detailed_results, } def to_json(self): # TODO pass def _add_trace_data_to_aggregate(aggregate, key, element): # Only cares about element['isa'], element['offline_aggregates'], and # element['plan']['level'] if not key in aggregate: aggregate[key] = { 'isa' : element['isa'] } for datakey in element['offline_aggregates'].keys(): aggregate[key][datakey] = [] if element['isa'] == 'state': aggregate[key]['attributes'] = ['power'] else: # TODO do not hardcode values aggregate[key]['attributes'] = ['duration', 'energy', 'rel_energy_prev', 'rel_energy_next'] if 'plan' in element and element['plan']['level'] == 'epilogue': aggregate[key]['attributes'].insert(0, 'timeout') attributes = aggregate[key]['attributes'].copy() for attribute in attributes: if attribute not in element['offline_aggregates']: aggregate[key]['attributes'].remove(attribute) for datakey, dataval in element['offline_aggregates'].items(): aggregate[key][datakey].extend(dataval) def pta_trace_to_aggregate(traces, ignore_trace_indexes = []): u""" Convert preprocessed DFA traces from peripherals/drivers to by_name aggregate for PTAModel. arguments: traces -- [ ... Liste von einzelnen Läufen (d.h. eine Zustands- und Transitionsfolge UNINITIALIZED -> foo -> FOO -> bar -> BAR -> ...) Jeder Lauf: - id: int Nummer des Laufs, beginnend bei 1 - trace: [ ... Liste von Zuständen und Transitionen Jeweils: - name: str Name - isa: str state // transition - parameter: { ... globaler Parameter: aktueller wert. null falls noch nicht eingestellt } - args: [ Funktionsargumente, falls isa == 'transition' ] - offline_aggregates: - power: [float(uW)] Mittlere Leistung während Zustand/Transitions - power_std: [float(uW^2)] Standardabweichung der Leistung - duration: [int(us)] Dauer - energy: [float(pJ)] Energieaufnahme des Zustands / der Transition - clip_rate: [float(0..1)] Clipping - paramkeys: [[str]] Name der berücksichtigten Parameter - param: [int // str] Parameterwerte. Quasi-Duplikat von 'parameter' oben Falls isa == 'transition': - timeout: [int(us)] Dauer des vorherigen Zustands - rel_energy_prev: [int(pJ)] - rel_energy_next: [int(pJ)] ] ] ignore_trace_indexes -- list of trace indexes. The corresponding taces will be ignored. returns a tuple of three elements: by_name -- measurements aggregated by state/transition name, annotated with parameter values parameter_names -- list of parameter names arg_count -- dict mapping transition names to the number of arguments of their corresponding driver function by_name layout: Dictionary with one key per state/transition ('send', 'TX', ...). Each element is in turn a dict with the following elements: - isa: 'state' or 'transition' - power: list of mean power measurements in µW - duration: list of durations in µs - power_std: list of stddev of power per state/transition - energy: consumed energy (power*duration) in pJ - paramkeys: list of parameter names in each measurement (-> list of lists) - param: list of parameter values in each measurement (-> list of lists) - attributes: list of keys that should be analyzed, e.g. ['power', 'duration'] additionally, only if isa == 'transition': - timeout: list of duration of previous state in µs - rel_energy_prev: transition energy relative to previous state mean power in pJ - rel_energy_next: transition energy relative to next state mean power in pJ """ arg_count = dict() by_name = dict() parameter_names = sorted(traces[0]['trace'][0]['parameter'].keys()) for run in traces: if run['id'] not in ignore_trace_indexes: for elem in run['trace']: if elem['isa'] == 'transition' and not elem['name'] in arg_count and 'args' in elem: arg_count[elem['name']] = len(elem['args']) if elem['name'] != 'UNINITIALIZED': _add_trace_data_to_aggregate(by_name, elem['name'], elem) for elem in by_name.values(): for key in elem['attributes']: elem[key] = np.array(elem[key]) return by_name, parameter_names, arg_count class PTAModel: u""" Parameter-aware PTA-based energy model. Supports both static and parameter-based model attributes, and automatic detection of parameter-dependence. The model heavily relies on two internal data structures: PTAModel.by_name and PTAModel.by_param. These provide measurements aggregated by state/transition name and (in case of by_para) parameter values. Layout: dictionary with one key per state/transition ('send', 'TX', ...) or one key per state/transition and parameter combination (('send', (1, 2)), ('send', (2, 3)), ('TX', (1, 2)), ('TX', (2, 3)), ...). For by_param, parameter values are ordered corresponding to the lexically sorted parameter names. Each element is in turn a dict with the following elements: - isa: 'state' or 'transition' - power: list of mean power measurements in µW - duration: list of durations in µs - power_std: list of stddev of power per state/transition - energy: consumed energy (power*duration) in pJ - paramkeys: list of parameter names in each measurement (-> list of lists) - param: list of parameter values in each measurement (-> list of lists) - attributes: list of keys that should be analyzed, e.g. ['power', 'duration'] additionally, only if isa == 'transition': - timeout: list of duration of previous state in µs - rel_energy_prev: transition energy relative to previous state mean power in pJ - rel_energy_next: transition energy relative to next state mean power in pJ """ def __init__(self, by_name, parameters, arg_count, traces = [], ignore_trace_indexes = [], discard_outliers = None, function_override = {}, verbose = True, use_corrcoef = False, hwmodel = None): """ Prepare a new PTA energy model. Actual model generation is done on-demand by calling the respective functions. arguments: by_name -- state/transition measurements aggregated by name, as returned by pta_trace_to_aggregate. parameters -- list of parameter names, as returned by pta_trace_to_aggregate arg_count -- function arguments, as returned by pta_trace_to_aggregate traces -- list of preprocessed DFA traces, as returned by RawData.get_preprocessed_data() ignore_trace_indexes -- list of trace indexes. The corresponding traces will be ignored. discard_outliers -- currently not supported: threshold for outlier detection and removel (float). Outlier detection is performed individually for each state/transition in each trace, so it only works if the benchmark ran several times. Given "data" (a set of measurements of the same thing, e.g. TX duration in the third benchmark trace), "m" (the median of all attribute measurements with the same parameters, which may include data from other traces), a data point X is considered an outlier if | 0.6745 * (X - m) / median(|data - m|) | > discard_outliers . function_override -- dict of overrides for automatic parameter function generation. If (state or transition name, model attribute) is present in function_override, the corresponding text string is the function used for analytic (parameter-aware/fitted) modeling of this attribute. It is passed to AnalyticFunction, see there for the required format. Note that this happens regardless of parameter dependency detection: The provided analytic function will be assigned even if it seems like the model attribute is static / parameter-independent. verbose -- print informative output, e.g. when removing an outlier use_corrcoef -- use correlation coefficient instead of stddev comparison to detect whether a model attribute depends on a parameter hwmodel -- hardware model suitable for PTA.from_hwmodel """ self.by_name = by_name self.by_param = by_name_to_by_param(by_name) self._parameter_names = sorted(parameters) self._num_args = arg_count self._use_corrcoef = use_corrcoef self.traces = traces self.stats = ParamStats(self.by_name, self.by_param, self._parameter_names, self._num_args, self._use_corrcoef, verbose = verbose) self.cache = {} np.seterr('raise') self._outlier_threshold = discard_outliers self.function_override = function_override self.verbose = verbose self.hwmodel = hwmodel self.ignore_trace_indexes = ignore_trace_indexes self._aggregate_to_ndarray(self.by_name) def distinct_param_values(self, state_or_tran, param_index = None, arg_index = None): if param_index != None: param_values = map(lambda x: x[param_index], self.by_name[state_or_tran]['param']) return sorted(set(param_values)) def _aggregate_to_ndarray(self, aggregate): for elem in aggregate.values(): for key in elem['attributes']: elem[key] = np.array(elem[key]) # This heuristic is very similar to the "function is not much better than # median" checks in get_fitted. So far, doing it here as well is mostly # a performance and not an algorithm quality decision. # --df, 2018-04-18 def depends_on_param(self, state_or_trans, key, param): return self.stats.depends_on_param(state_or_trans, key, param) # See notes on depends_on_param def depends_on_arg(self, state_or_trans, key, param): return self.stats.depends_on_arg(state_or_trans, key, param) def _get_model_from_dict(self, model_dict, model_function): model = {} for name, elem in model_dict.items(): model[name] = {} for key in elem['attributes']: try: model[name][key] = model_function(elem[key]) except RuntimeWarning: vprint(self.verbose, '[W] Got no data for {} {}'.format(name, key)) except FloatingPointError as fpe: vprint(self.verbose, '[W] Got no data for {} {}: {}'.format(name, key, fpe)) return model def get_static(self): """ Get static model function: name, attribute -> model value. Uses the median of by_name for modeling. """ static_model = self._get_model_from_dict(self.by_name, np.median) def static_median_getter(name, key, **kwargs): return static_model[name][key] return static_median_getter def get_static_using_mean(self): """ Get static model function: name, attribute -> model value. Uses the mean of by_name for modeling. """ static_model = self._get_model_from_dict(self.by_name, np.mean) def static_mean_getter(name, key, **kwargs): return static_model[name][key] return static_mean_getter def get_param_lut(self, fallback = False): """ Get parameter-look-up-table model function: name, attribute, parameter values -> model value. The function can only give model values for parameter combinations present in by_param. By default, it raises KeyError for other values. arguments: fallback -- Fall back to the (non-parameter-aware) static model when encountering unknown parameter values """ static_model = self._get_model_from_dict(self.by_name, np.median) lut_model = self._get_model_from_dict(self.by_param, np.median) def lut_median_getter(name, key, param, arg = [], **kwargs): param.extend(map(soft_cast_int, arg)) try: return lut_model[(name, tuple(param))][key] except KeyError: if fallback: return static_model[name][key] raise return lut_median_getter def param_index(self, param_name): if param_name in self._parameter_names: return self._parameter_names.index(param_name) return len(self._parameter_names) + int(param_name) def param_name(self, param_index): if param_index < len(self._parameter_names): return self._parameter_names[param_index] return str(param_index) def get_fitted(self, safe_functions_enabled = False): """ Get paramete-aware model function and model information function. Returns two functions: model_function(name, attribute, param=parameter values) -> model value. model_info(name, attribute) -> {'fit_result' : ..., 'function' : ... } or None """ if 'fitted_model_getter' in self.cache and 'fitted_info_getter' in self.cache: return self.cache['fitted_model_getter'], self.cache['fitted_info_getter'] static_model = self._get_model_from_dict(self.by_name, np.median) param_model = dict([[state_or_tran, {}] for state_or_tran in self.by_name.keys()]) paramfit = ParallelParamFit(self.by_param) for state_or_tran in self.by_name.keys(): for model_attribute in self.by_name[state_or_tran]['attributes']: fit_results = {} for parameter_index, parameter_name in enumerate(self._parameter_names): if self.depends_on_param(state_or_tran, model_attribute, parameter_name): paramfit.enqueue(state_or_tran, model_attribute, parameter_index, parameter_name, safe_functions_enabled) if arg_support_enabled and self.by_name[state_or_tran]['isa'] == 'transition': for arg_index in range(self._num_args[state_or_tran]): if self.depends_on_arg(state_or_tran, model_attribute, arg_index): paramfit.enqueue(state_or_tran, model_attribute, len(self._parameter_names) + arg_index, arg_index, safe_functions_enabled) paramfit.fit() for state_or_tran in self.by_name.keys(): num_args = 0 if arg_support_enabled and self.by_name[state_or_tran]['isa'] == 'transition': num_args = self._num_args[state_or_tran] for model_attribute in self.by_name[state_or_tran]['attributes']: fit_results = {} for result in paramfit.results: if result['key'][0] == state_or_tran and result['key'][1] == model_attribute: fit_result = result['result'] if fit_result['best_rmsd'] >= min(fit_result['mean_rmsd'], fit_result['median_rmsd']): vprint(self.verbose, '[I] Not modeling {} {} as function of {}: best ({:.0f}) is worse than ref ({:.0f}, {:.0f})'.format( state_or_tran, model_attribute, result['key'][2], fit_result['best_rmsd'], fit_result['mean_rmsd'], fit_result['median_rmsd'])) # See notes on depends_on_param elif fit_result['best_rmsd'] >= 0.8 * min(fit_result['mean_rmsd'], fit_result['median_rmsd']): vprint(self.verbose, '[I] Not modeling {} {} as function of {}: best ({:.0f}) is not much better than ({:.0f}, {:.0f})'.format( state_or_tran, model_attribute, result['key'][2], fit_result['best_rmsd'], fit_result['mean_rmsd'], fit_result['median_rmsd'])) else: fit_results[result['key'][2]] = fit_result if (state_or_tran, model_attribute) in self.function_override: function_str = self.function_override[(state_or_tran, model_attribute)] x = AnalyticFunction(function_str, self._parameter_names, num_args) x.fit(self.by_param, state_or_tran, model_attribute) if x.fit_success: param_model[state_or_tran][model_attribute] = { 'fit_result': fit_results, 'function' : x } elif len(fit_results.keys()): x = analytic.function_powerset(fit_results, self._parameter_names, num_args) x.fit(self.by_param, state_or_tran, model_attribute) if x.fit_success: param_model[state_or_tran][model_attribute] = { 'fit_result': fit_results, 'function' : x } def model_getter(name, key, **kwargs): if 'arg' in kwargs and 'param' in kwargs: kwargs['param'].extend(map(soft_cast_int, kwargs['arg'])) if key in param_model[name]: param_list = kwargs['param'] param_function = param_model[name][key]['function'] if param_function.is_predictable(param_list): return param_function.eval(param_list) return static_model[name][key] def info_getter(name, key): if key in param_model[name]: return param_model[name][key] return None self.cache['fitted_model_getter'] = model_getter self.cache['fitted_info_getter'] = info_getter return model_getter, info_getter def to_json(self): static_model = self.get_static() _, param_info = self.get_fitted() pta = PTA.from_json(self.hwmodel) pta.update(static_model, param_info) return pta.to_json() def states(self): return sorted(list(filter(lambda k: self.by_name[k]['isa'] == 'state', self.by_name.keys()))) def transitions(self): return sorted(list(filter(lambda k: self.by_name[k]['isa'] == 'transition', self.by_name.keys()))) def states_and_transitions(self): ret = self.states() ret.extend(self.transitions()) return ret def parameters(self): return self._parameter_names def attributes(self, state_or_trans): return self.by_name[state_or_trans]['attributes'] def assess(self, model_function): """ Calculate MAE, SMAPE, etc. of model_function for each by_name entry. state/transition/... name and parameter values are fed into model_function. The by_name entries of this PTAModel are used as ground truth and compared with the values predicted by model_function. If 'traces' was set when creating this object, the model quality is also assessed on a per-trace basis. For proper model assessments, the data used to generate model_function and the data fed into this AnalyticModel instance must be mutually exclusive (e.g. by performing cross validation). Otherwise, overfitting cannot be detected. """ detailed_results = {} model_energy_list = [] real_energy_list = [] model_rel_energy_list = [] model_state_energy_list = [] model_duration_list = [] real_duration_list = [] model_timeout_list = [] real_timeout_list = [] for name, elem in sorted(self.by_name.items()): detailed_results[name] = {} for key in elem['attributes']: predicted_data = np.array(list(map(lambda i: model_function(name, key, param=elem['param'][i]), range(len(elem[key]))))) measures = regression_measures(predicted_data, elem[key]) detailed_results[name][key] = measures for trace in self.traces: if trace['id'] not in self.ignore_trace_indexes: for rep_id in range(len(trace['trace'][0]['offline'])): model_energy = 0. real_energy = 0. model_rel_energy = 0. model_state_energy = 0. model_duration = 0. real_duration = 0. model_timeout = 0. real_timeout = 0. for i, trace_part in enumerate(trace['trace']): name = trace_part['name'] prev_name = trace['trace'][i-1]['name'] isa = trace_part['isa'] if name != 'UNINITIALIZED': param = trace_part['offline_aggregates']['param'][rep_id] prev_param = trace['trace'][i-1]['offline_aggregates']['param'][rep_id] power = trace_part['offline'][rep_id]['uW_mean'] duration = trace_part['offline'][rep_id]['us'] prev_duration = trace['trace'][i-1]['offline'][rep_id]['us'] real_energy += power * duration if isa == 'state': model_energy += model_function(name, 'power', param=param) * duration else: model_energy += model_function(name, 'energy', param=param) # If i == 1, the previous state was UNINITIALIZED, for which we do not have model data if i == 1: model_rel_energy += model_function(name, 'energy', param=param) else: model_rel_energy += model_function(prev_name, 'power', param=prev_param) * (prev_duration + duration) model_state_energy += model_function(prev_name, 'power', param=prev_param) * (prev_duration + duration) model_rel_energy += model_function(name, 'rel_energy_prev', param=param) real_duration += duration model_duration += model_function(name, 'duration', param=param) if 'plan' in trace_part and trace_part['plan']['level'] == 'epilogue': real_timeout += trace_part['offline'][rep_id]['timeout'] model_timeout += model_function(name, 'timeout', param=param) real_energy_list.append(real_energy) model_energy_list.append(model_energy) model_rel_energy_list.append(model_rel_energy) model_state_energy_list.append(model_state_energy) real_duration_list.append(real_duration) model_duration_list.append(model_duration) real_timeout_list.append(real_timeout) model_timeout_list.append(model_timeout) if len(self.traces): return { 'by_name' : detailed_results, 'duration_by_trace' : regression_measures(np.array(model_duration_list), np.array(real_duration_list)), 'energy_by_trace' : regression_measures(np.array(model_energy_list), np.array(real_energy_list)), 'timeout_by_trace' : regression_measures(np.array(model_timeout_list), np.array(real_timeout_list)), 'rel_energy_by_trace' : regression_measures(np.array(model_rel_energy_list), np.array(real_energy_list)), 'state_energy_by_trace' : regression_measures(np.array(model_state_energy_list), np.array(real_energy_list)), } return { 'by_name' : detailed_results } class MIMOSA: """ MIMOSA log loader for DFA traces with auto-calibration. Expects a MIMOSA log file generated via dfatool and a dfatool-generated benchmark: There is an automatic calibration step at the start and the trigger pin is high iff a transition is active. The resulting data is a list of state/transition/state/transition/... measurements. """ def __init__(self, voltage, shunt, verbose = True): """ Initialize MIMOSA loader for a specific voltage and shunt setting. arguments: voltage -- MIMOSA voltage used for measurements shunt -- Shunt value in Ohms verbose -- notify about invalid data and the likes """ self.voltage = voltage self.shunt = shunt self.verbose = verbose self.r1 = 984 # "1k" self.r2 = 99013 # "100k" def charge_to_current_nocal(self, charge): u"""Convert charge per 10µs to mean currents without accounting for calibration.""" ua_max = 1.836 / self.shunt * 1000000 ua_step = ua_max / 65535 return charge * ua_step def _load_tf(self, tf): num_bytes = tf.getmember('/tmp/mimosa//mimosa_scale_1.tmp').size charges = np.ndarray(shape=(int(num_bytes / 4)), dtype=np.int32) triggers = np.ndarray(shape=(int(num_bytes / 4)), dtype=np.int8) with tf.extractfile('/tmp/mimosa//mimosa_scale_1.tmp') as f: content = f.read() iterator = struct.iter_unpack('> 4) triggers[i] = (word[0] & 0x08) >> 3 i += 1 return charges, triggers def load_data(self, raw_data): """Load a MIMOSA log archive from a raw bytestring.""" with io.BytesIO(raw_data) as data_object: with tarfile.open(fileobj = data_object) as tf: return self._load_tf(tf) def load_file(self, filename): """Load a MIMOSA log archive from a filename.""" with tarfile.open(filename) as tf: return self._load_tf(tf) def currents_nocal(self, charges): u"""Convert charge per 10µs to mean currents without accounting for calibration.""" ua_max = 1.836 / self.shunt * 1000000 ua_step = ua_max / 65535 return charges.astype(np.double) * ua_step def trigger_edges(self, triggers): """ Return indexes of trigger edges (both 0->1 and 1->0) in log data. arguments: triggers -- trigger array as returned by load_data Ignores the first 10 seconds, which are used for calibration and may contain bogus triggers due to DUT resets. Returns a list of int. """ trigidx = [] prevtrig = triggers[0] # the device is reset for MIMOSA calibration in the first 10s and may # send bogus interrupts -> bogus triggers for i in range(1000000, triggers.shape[0]): trig = triggers[i] if trig != prevtrig: # Due to MIMOSA's integrate-read-reset cycle, the trigger # appears two points (20µs) before the corresponding data trigidx.append(i+2) prevtrig = trig return trigidx def calibration_edges(self, currents): """ Return start/stop indexes of calibration measurements. arguments: currents -- uncalibrated currents as reported by MIMOSA. For best results, it may help to use a running mean, like so: currents = running_mean(currents_nocal(..., 10)) Returns six indexes: - Disconnected start - Disconnected stop - R1 (1 kOhm) start - R1 (1 kOhm) stop - R2 (100 kOhm) start - R2 (100 kOhm) stop """ r1idx = 0 r2idx = 0 ua_r1 = self.voltage / self.r1 * 1000000 # first second may be bogus for i in range(100000, len(currents)): if r1idx == 0 and currents[i] > ua_r1 * 0.6: r1idx = i elif r1idx != 0 and r2idx == 0 and i > (r1idx + 180000) and currents[i] < ua_r1 * 0.4: r2idx = i # 2s disconnected, 2s r1, 2s r2 with r1 < r2 -> ua_r1 > ua_r2 # allow 5ms buffer in both directions to account for bouncing relais contacts return r1idx - 180500, r1idx - 500, r1idx + 500, r2idx - 500, r2idx + 500, r2idx + 180500 def calibration_function(self, charges, cal_edges): u""" Calculate calibration function from previously determined calibration phase. arguments: charges -- raw charges from MIMOSA cal_edges -- calibration edges as returned by calibration_edges returns (calibration_function, calibration_data): calibration_function -- charge in pJ (float) -> current in uA (float). Converts the amount of charge in a 10 µs interval to the mean current during the same interval. calibration_data -- dict containing the following keys: edges -- calibration points in the log file, in µs offset -- ... offset2 -- ... slope_low -- ... slope_high -- ... add_low -- ... add_high -- .. r0_err_uW -- mean error of uncalibrated data at "∞ Ohm" in µW r0_std_uW -- standard deviation of uncalibrated data at "∞ Ohm" in µW r1_err_uW -- mean error of uncalibrated data at 1 kOhm r1_std_uW -- stddev at 1 kOhm r2_err_uW -- mean error at 100 kOhm r2_std_uW -- stddev at 100 kOhm """ dis_start, dis_end, r1_start, r1_end, r2_start, r2_end = cal_edges if dis_start < 0: dis_start = 0 chg_r0 = charges[dis_start:dis_end] chg_r1 = charges[r1_start:r1_end] chg_r2 = charges[r2_start:r2_end] cal_0_mean = np.mean(chg_r0) cal_r1_mean = np.mean(chg_r1) cal_r2_mean = np.mean(chg_r2) ua_r1 = self.voltage / self.r1 * 1000000 ua_r2 = self.voltage / self.r2 * 1000000 if cal_r2_mean > cal_0_mean: b_lower = (ua_r2 - 0) / (cal_r2_mean - cal_0_mean) else: vprint(self.verbose, '[W] 0 uA == %.f uA during calibration' % (ua_r2)) b_lower = 0 b_upper = (ua_r1 - ua_r2) / (cal_r1_mean - cal_r2_mean) a_lower = -b_lower * cal_0_mean a_upper = -b_upper * cal_r2_mean if self.shunt == 680: # R1 current is higher than shunt range -> only use R2 for calibration def calfunc(charge): if charge < cal_0_mean: return 0 else: return charge * b_lower + a_lower else: def calfunc(charge): if charge < cal_0_mean: return 0 if charge <= cal_r2_mean: return charge * b_lower + a_lower else: return charge * b_upper + a_upper + ua_r2 caldata = { 'edges' : [x * 10 for x in cal_edges], 'offset': cal_0_mean, 'offset2' : cal_r2_mean, 'slope_low' : b_lower, 'slope_high' : b_upper, 'add_low' : a_lower, 'add_high' : a_upper, 'r0_err_uW' : np.mean(self.currents_nocal(chg_r0)) * self.voltage, 'r0_std_uW' : np.std(self.currents_nocal(chg_r0)) * self.voltage, 'r1_err_uW' : (np.mean(self.currents_nocal(chg_r1)) - ua_r1) * self.voltage, 'r1_std_uW' : np.std(self.currents_nocal(chg_r1)) * self.voltage, 'r2_err_uW' : (np.mean(self.currents_nocal(chg_r2)) - ua_r2) * self.voltage, 'r2_std_uW' : np.std(self.currents_nocal(chg_r2)) * self.voltage, } #print("if charge < %f : return 0" % cal_0_mean) #print("if charge <= %f : return charge * %f + %f" % (cal_r2_mean, b_lower, a_lower)) #print("else : return charge * %f + %f + %f" % (b_upper, a_upper, ua_r2)) return calfunc, caldata """ def calcgrad(self, currents, threshold): grad = np.gradient(running_mean(currents * self.voltage, 10)) # len(grad) == len(currents) - 9 subst = [] lastgrad = 0 for i in range(len(grad)): # minimum substate duration: 10ms if np.abs(grad[i]) > threshold and i - lastgrad > 50: # account for skew introduced by running_mean and current # ramp slope (parasitic capacitors etc.) subst.append(i+10) lastgrad = i if lastgrad != i: subst.append(i+10) return subst # TODO konfigurierbare min/max threshold und len(gradidx) > X, binaere # Sache nach noetiger threshold. postprocessing mit # "zwei benachbarte substates haben sehr aehnliche werte / niedrige stddev" -> mergen # ... min/max muessen nicht vorgegeben werden, sind ja bekannt (0 / np.max(grad)) # TODO bei substates / index foo den offset durch running_mean beachten # TODO ggf. clustering der 'abs(grad) > threshold' und bestimmung interessanter # uebergaenge dadurch? def gradfoo(self, currents): gradients = np.abs(np.gradient(running_mean(currents * self.voltage, 10))) gradmin = np.min(gradients) gradmax = np.max(gradients) threshold = np.mean([gradmin, gradmax]) gradidx = self.calcgrad(currents, threshold) num_substates = 2 while len(gradidx) != num_substates: if gradmax - gradmin < 0.1: # We did our best return threshold, gradidx if len(gradidx) > num_substates: gradmin = threshold else: gradmax = threshold threshold = np.mean([gradmin, gradmax]) gradidx = self.calcgrad(currents, threshold) return threshold, gradidx """ def analyze_states(self, charges, trigidx, ua_func): u""" Split log data into states and transitions and return mean power and duration for each element. arguments: charges -- raw charges (each element describes the charge transferred during 10 µs) trigidx -- "charges" indexes corresponding to a trigger edge ua_func -- charge -> current function as returned by calibration_function returns a list of (alternating) states and transitions. Each element is a dict containing: - isa: 'state' oder 'transition' - clip_rate: range(0..1) Anteil an Clipping im Energieverbrauch - raw_mean: Mittelwert der Rohwerte - raw_std: Standardabweichung der Rohwerte - uW_mean: Mittelwert der (kalibrierten) Leistungsaufnahme - uW_std: Standardabweichung der (kalibrierten) Leistungsaufnahme - us: Dauer Nur falls isa == 'transition': - timeout: Dauer des vorherigen Zustands - uW_mean_delta_prev: Differenz zwischen uW_mean und uW_mean des vorherigen Zustands - uW_mean_delta_next: Differenz zwischen uW_mean und uW_mean des Folgezustands """ previdx = 0 is_state = True iterdata = [] for idx in trigidx: range_raw = charges[previdx:idx] range_ua = ua_func(range_raw) substates = {} if previdx != 0 and idx - previdx > 200: thr, subst = 0, [] #self.gradfoo(range_ua) if len(subst): statelist = [] prevsubidx = 0 for subidx in subst: statelist.append({ 'duration': (subidx - prevsubidx) * 10, 'uW_mean' : np.mean(range_ua[prevsubidx : subidx] * self.voltage), 'uW_std' : np.std(range_ua[prevsubidx : subidx] * self.voltage), }) prevsubidx = subidx substates = { 'threshold' : thr, 'states' : statelist, } isa = 'state' if not is_state: isa = 'transition' data = { 'isa': isa, 'clip_rate' : np.mean(range_raw == 65535), 'raw_mean': np.mean(range_raw), 'raw_std' : np.std(range_raw), 'uW_mean' : np.mean(range_ua * self.voltage), 'uW_std' : np.std(range_ua * self.voltage), 'us' : (idx - previdx) * 10, } if 'states' in substates: data['substates'] = substates ssum = np.sum(list(map(lambda x : x['duration'], substates['states']))) if ssum != data['us']: vprint(self.verbose, "ERR: duration %d vs %d" % (data['us'], ssum)) if isa == 'transition': # subtract average power of previous state # (that is, the state from which this transition originates) data['uW_mean_delta_prev'] = data['uW_mean'] - iterdata[-1]['uW_mean'] # placeholder to avoid extra cases in the analysis data['uW_mean_delta_next'] = data['uW_mean'] data['timeout'] = iterdata[-1]['us'] elif len(iterdata) > 0: # subtract average power of next state # (the state into which this transition leads) iterdata[-1]['uW_mean_delta_next'] = iterdata[-1]['uW_mean'] - data['uW_mean'] iterdata.append(data) previdx = idx is_state = not is_state return iterdata