def _get_single_fragment_amplitudes(fragment, init_wires = None, exit_wires = None,
**kwargs):
'''
this method accepts:
(i) a circuit fragment,
(ii) a list of "input" wires (`init_wires`),
(iii) a list of "output" wires (`exit_wires`),
(iv) additional keyword arguments passed to the Qiskit statevector simulator.
this method returns "conditional amplitudes" of the fragment, which are
(i) conditional on computational basis states prepared at `init_wires`,
(ii) projected onto computational basis states at `exit_wires`.
'''
# initialize an empty conditional distribution over measurement outcomes
frag_amps = FragmentAmplitudes()
# identify computational basis states
def _state_ZXY(bit): return "+Z" if not bit else "-Z"
def _state_vec(bit):
vec = (1,0) if not bit else (0,1)
return np.array(vec, dtype = complex)
# identify the axes we will project out for the exit wires
# and sort exit wires according to these axes
exit_axes = { wire : -fragment.qubits.index(wire)-1 for wire in exit_wires }
exit_wires = sorted(exit_wires, key = lambda wire : exit_axes[wire] )
# sort init wires to fix their order
init_wires = sorted(init_wires, key = lambda wire : fragment.qubits.index(wire))
# for every choice of states prepared on all on init wires
for init_states in set_product(range(2), repeat = len(init_wires)):
init_conds = frozenset(zip(init_wires, init_states))
# build circuit with the "input" states prepared appropriately
prep_circuits = [ _act_gate(fragment, _prep_gate(_state_ZXY(bit_state)), wire)
for wire, bit_state in init_conds ]
init_circuit = reduce(lambda x, y : x + y, prep_circuits + [ fragment ])
# get the quantum state vector for the circuit
all_amplitudes = get_circuit_amplitudes(init_circuit, **kwargs)
# for every set of projections on all on exit wires
for exit_states in set_product(range(2), repeat = len(exit_wires)):
exit_conds = zip(exit_wires, exit_states)
# project onto the given measured states on exit wires
projected_amplitudes = deepcopy(all_amplitudes)
for exit_wire, exit_state in zip(exit_wires, exit_states):
exit_vec = tf.constant(_state_vec(exit_state))
axes = [ [ 0 ], [ exit_axes[exit_wire] ] ]
projected_amplitudes \
= tf.tensordot(exit_vec, projected_amplitudes, axes = axes)
frag_amps.add(init_conds, exit_conds, projected_amplitudes)
return frag_amps
def alternating_bubbler(lattice_shape, lattice_vectors, link_tensors=False):
if not link_tensors:
base_vec, other_vecs = lattice_vectors[0], lattice_vectors[1:]
signed_ones = set_product([+1, -1], repeat=len(lattice_shape) - 1)
new_vectors = [
_add(base_vec, sign * np.array(vec)) for signs in signed_ones
for sign, vec in zip(signs, other_vecs)
]
extended_lattice_shape = [
dim if dim % 2 == 0 else dim + 1 for dim in lattice_shape
]
fst_scan = scanning_bubbler(extended_lattice_shape, new_vectors)
snd_scan = [
_add(pos, base_vec, extended_lattice_shape) for pos in fst_scan
]
full_scan = sorted(fst_scan) + sorted(snd_scan)
return [
pos for pos in full_scan
if all(pp < ss for pp, ss in zip(pos, lattice_shape))
]
else:
zero_pos = (0, ) * len(lattice_shape)
half_vecs = [_half(vec) for vec in lattice_vectors]
sublattice_scans = [
_crystal_sites(lattice_shape, lattice_vectors, initial_pos)
for initial_pos in half_vecs + [zero_pos]
]
return [
pos for sublattice_scan in sublattice_scans
for pos in sublattice_scan
]
def query_united_distribution(frag_dists,
wire_path_map,
circuit_wires,
frag_wires,
query_states=[],
force_probs=True):
# identify all cuts ("stitches") with a dictionary mapping exit wires to init wires
stitches = _identify_stitches(wire_path_map, circuit_wires)
# determine metadata for the united distribution
frag_metadata = _get_distribution_metadata(frag_dists)
_, dist_obj_type, dist_dat_type = frag_metadata
# figure out how to query distributions
if dist_obj_type is tf.SparseTensor:
def _query(dist, query_state):
value = 0
for idx, state in enumerate(dist.indices.numpy()):
if all(np.equal(state, query_state)):
value = dist.values.numpy()[idx]
break
return value
else:
def _query(dist, state):
return dist.numpy()[state]
# identify the operators and scalar factor at each stitch
_, stitch_ops, scalar_factor \
= _get_uniting_objects(frag_dists, stitches, frag_metadata)
# initialize the values we are querying,
# and identify the state on each fragment for each query state
vals_type = complex if _is_complex(dist_dat_type) else float
state_vals = np.zeros(len(query_states), dtype=vals_type)
frag_query_states = get_frag_states(query_states, wire_path_map,
circuit_wires, frag_wires)
# loop over all assigments of stitch operators at all cut locations (stitches)
for op_assignment in set_product(stitch_ops, repeat=len(stitches)):
# collect tensor factors of this term in the combined distribution
# and add to the values we are querying
dist_factors = _collect_tensor_factors(frag_dists, stitches,
op_assignment)
for query_idx, frag_states in enumerate(frag_query_states):
dist_state_iter = zip(dist_factors, frag_states)
state_vals[query_idx] \
+= np.prod([ _query(dist,state) for dist, state in dist_state_iter ])
# convert amplitudes to probabilities if appropriate
if force_probs and _is_complex(dist_dat_type):
state_vals = abs(state_vals)**2
return scalar_factor * state_vals
def shuffle_bases(self, init_basis=None, exit_basis=None):
new_amps = FragmentAmplitudes()
# identify wires with init/exit conditions
init_wires = self.init_wires()
exit_wires = self.exit_wires()
# set default bases
if init_basis == None: init_basis = {wire: 0 for wire in init_wires}
if exit_basis == None: exit_basis = {wire: 0 for wire in exit_wires}
# return a given qubit state in a basis in which zero_state is |0>
def _relative_state(wire, state, zero_state):
if state == 0:
return zero_state
else:
if zero_state in (0, 1):
return 1 - zero_state
else:
return tuple(-xx for xx in zero_state)
# loop over all assignments of init/exit conditions in the appropriate bases
for init_states in set_product([0, 1], repeat=len(init_wires)):
new_init_conds = list(zip(init_wires, init_states))
old_init_conds = {
wire: _relative_state(wire, state, init_basis[wire])
for wire, state in new_init_conds
}
for exit_states in set_product([0, 1], repeat=len(exit_wires)):
new_exit_conds = list(zip(exit_wires, exit_states))
old_exit_conds = {
wire: _relative_state(wire, state, exit_basis[wire])
for wire, state in new_exit_conds
}
new_amps.add(new_init_conds, new_exit_conds,
self[old_init_conds, old_exit_conds])
return new_amps
def shuffle_bases(self, init_basis=pauli, exit_basis=pauli):
new_probs = FragmentProbabilities(init_basis, exit_basis)
# identify wires with init/exit conditions
init_wires = self.init_wires()
exit_wires = self.exit_wires()
if type(init_basis) is str:
init_basis = basis_ops[init_basis]
if type(exit_basis) is str:
exit_basis = basis_ops[exit_basis]
# loop over all assignments of init/exit conditions in the appropriate bases
for init_ops in set_product(init_basis, repeat=len(init_wires)):
new_init_conds = list(zip(init_wires, init_ops))
for exit_ops in set_product(exit_basis, repeat=len(exit_wires)):
new_exit_conds = list(zip(exit_wires, exit_ops))
new_probs.add(new_init_conds, new_exit_conds,
self[new_init_conds, new_exit_conds])
return new_probs
def tf_outer_product(tensor_a, tensor_b):
if type(tensor_a) is not tf.SparseTensor:
return tf.tensordot(tensor_a, tensor_b, axes=0)
else:
index_iterator = set_product(tensor_a.indices.numpy(),
tensor_b.indices.numpy())
indices = [np.concatenate(index_pair) for index_pair in index_iterator]
values_a = tensor_a.values.numpy()
values_b = tensor_b.values.numpy()
values_shape = (len(tensor_a.values), len(tensor_b.values))
values = np.empty(values_shape, dtype=values_a.dtype)
values = np.outer(values_a, values_b, values).flatten()
dense_shape = tf.concat([tensor_a.dense_shape, tensor_b.dense_shape],
0)
return tf.SparseTensor(indices, values, dense_shape)
def vertex_tensor_XY(dimension, bond_dimension, inv_temp, field):
assert (bond_dimension % 2 == 1) # only allow for odd bond dimensions
def _prod_diag_val(indices, xx):
return np.prod([scipy.special.iv(idx, xx) for idx in indices])
def _mod_diag_val(indices, xx):
idx_sum = sum(indices[:dimension]) - sum(indices[dimension:])
return scipy.special.iv(idx_sum, xx)
index_vals = set_product(_integers(bond_dimension, center_on_zero=True),
repeat=2 * dimension)
vector = tf.constant([
np.sqrt(_prod_diag_val(indices, inv_temp)) *
_mod_diag_val(indices, inv_temp * field) for indices in index_vals
])
return tf.reshape(vector, [bond_dimension] * 2 * dimension)
def unite_fragment_distributions(frag_dists,
wire_path_map,
circuit_wires,
frag_wires,
force_probs=True,
status_updates=False):
# identify all cuts ("stitches") with a dictionary mapping exit wires to init wires
stitches = _identify_stitches(wire_path_map, circuit_wires)
# determine metadata for the united distribution
frag_metadata = _get_distribution_metadata(frag_dists)
_, _, dist_dat_type = frag_metadata
# initialize an empty distribution
# and identify the operators / scalar factor at each stitch
united_dist, stitch_ops, scalar_factor \
= _get_uniting_objects(frag_dists, stitches, frag_metadata)
# pre-process distributions to switch conditions into the pauli basis
if stitch_ops == basis_ops_pauli:
frag_dists = [
frag_dist if frag_dist.init_basis == pauli
and frag_dist.exit_basis == pauli else frag_dist.shuffle_bases(
pauli, pauli) for frag_dist in frag_dists
]
# loop over all assigments of stitch operators at all cut locations (stitches)
for op_assignment in set_product(stitch_ops, repeat=len(stitches)):
if status_updates: print(op_assignment)
# collect tensor factors of this term in the combined distribution
# and add to the united distribution
dist_factors = _collect_tensor_factors(frag_dists, stitches,
op_assignment)
united_dist += reduce(tf_outer_product, dist_factors[::-1])
if status_updates: print()
# convert amplitudes to probabilities if appropriate
if force_probs and _is_complex(dist_dat_type):
united_dist = abs(united_dist)**2
# sort wires/qubits appropriately before returning the distribution
perm = _united_axis_permutation(wire_path_map, circuit_wires, frag_wires)
return scalar_factor * tf_transpose(united_dist, perm)
def checkerboard_tensor(dimension, spokes, inv_temp, field):
tensor_legs = 2**dimension
# shift a vertex on a hypercube to an adjacent vertex in a given direction
# all vertices are labeled by bitstrings (represented by integers),
# so moving in a particular direction corresponds to XORing with a bitstring
# that is only `1` on the bit corresponding to the given direction
def _shift(idx, direction):
return idx ^ (2**direction)
def _angles_coeff(angles):
site_term = field * sum(np.cos(angles))
link_term = sum(
np.cos(angles[idx] - angles[_shift(idx, direction)])
for idx in range(tensor_legs) for direction in range(dimension))
return np.exp(inv_temp / 2 * (site_term + link_term))
all_angles = set_product(_angles(spokes), repeat=tensor_legs)
vector = tf.constant([_angles_coeff(angles) for angles in all_angles])
return tf.reshape(vector,
(spokes, ) * tensor_legs) / spokes**(tensor_legs / 2)
def to_probabilities(self,
init_basis=ZZXY,
exit_basis=ZZXY,
dtype=tf.float64):
# save the bases in which we want to store init/exit conditions
final_init_basis = init_basis
final_exit_basis = exit_basis
# we can only actually *compute* distributions in certain bases,
# so if we weren't asked for one of those, choose one of them to use for now
if init_basis not in [SIC, ZZXY]:
init_basis = ZZXY
if exit_basis not in [SIC, ZZXY]:
exit_basis = ZZXY
# identify computational basis states and coefficients for each SIC / ZZXY state
def _dist_terms_SIC(oper, conjugate):
assert (oper in basis_ops_SIC)
sign = 1 if not conjugate else -1
theta, phi = get_bloch_angles(state_vecs_SIC[oper])
# | theta, phi > = cos(theta/2) |0> + exp(i phi) sin(theta/2) | 1 >
return [(0, np.cos(theta / 2)),
(1, np.exp(sign * 1j * phi) * np.sin(theta / 2))]
def _dist_terms_ZZXY(oper, conjugate):
assert (oper in basis_ops_ZZXY)
if oper == "-Z": # | -Z > = 1 | 1 >
return [(1, 1)]
if oper == "+Z": # | +Z > = 1 | 0 >
return [(0, 1)]
if oper == "+X": # | +X > = ( | 0 > + | 1 > ) / sqrt(2)
return [(0, 1 / np.sqrt(2)), (1, 1 / np.sqrt(2))]
if oper == "+Y": # | +Y > = ( | 0 > + i | 1 > ) / sqrt(2)
sign = 1 if not conjugate else -1
return [(0, 1 / np.sqrt(2)), (1, sign * 1j / np.sqrt(2))]
_dist_terms = {SIC: _dist_terms_SIC, ZZXY: _dist_terms_ZZXY}
# determine which basis of operators to use for init/exit conditions,
# as well as corresponding computational basis states / coefficients
init_basis_ops = basis_ops[init_basis]
exit_basis_ops = basis_ops[exit_basis]
_init_dist_terms = _dist_terms[init_basis]
_exit_dist_terms = _dist_terms[exit_basis]
# identify wires with init/exit conditions
init_wires = self.init_wires()
exit_wires = self.exit_wires()
# initialize a conditional probability distribution
probs = FragmentProbabilities(init_basis, exit_basis)
# loop over all init/exit conditions (init_states/exit_states)
for init_states in set_product(init_basis_ops, repeat=len(init_wires)):
# conditions (for probs) corresponding to this choice of init_states
prob_init_conds = {(True, wire, state)
for wire, state in zip(init_wires, init_states)}
# computational basis terms that contribute to this choice of init_states
init_terms = [
_init_dist_terms(state, False) for state in init_states
]
for exit_states in set_product(exit_basis_ops,
repeat=len(exit_wires)):
prob_exit_conds = {
(False, wire, state)
for wire, state in zip(exit_wires, exit_states)
}
exit_terms = [
_exit_dist_terms(state, True) for state in exit_states
]
state_amps = 0 # empty vector of amplitudes for these init/exit_states
# looping over all contributing terms to this choice of init/exit_states
for init_bits_facs in set_product(*init_terms):
try:
init_bits, init_facs = zip(*init_bits_facs)
except:
init_bits, init_facs = [], []
# scalar factor associated with this set of terms
init_fac = np.prod(init_facs)
# conditions (i.e. on the amplitude distribution) for to this term
amp_init_conds = {
(True, wire, bit)
for wire, bit in zip(init_wires, init_bits)
}
for exit_bits_facs in set_product(*exit_terms):
try:
exit_bits, exit_facs = zip(*exit_bits_facs)
except:
exit_bits, exit_facs = [], []
exit_fac = np.prod(exit_facs)
amp_exit_conds = {
(False, wire, bit)
for wire, bit in zip(exit_wires, exit_bits)
}
# add to the amplitudes for this choice of init/exit_states
fac = init_fac * exit_fac
state_amps += fac * self[amp_init_conds,
amp_exit_conds]
# having collected amplitudes, convert them to probabilities
cond_probs = tf.cast(abs(state_amps)**2, dtype=dtype)
probs.add(prob_init_conds, prob_exit_conds, cond_probs)
# if we computed distributions in the same init/exit bases as we were asked for,
# then just return the conditional distribution we computed
# otherwise, change bases appropriately
if init_basis == final_init_basis and exit_basis == final_exit_basis:
return probs
else:
return probs.shuffle_bases(final_init_basis, final_exit_basis)
def sample_positive_distribution(frag_dists,
wire_path_map,
circuit_wires,
frag_wires,
num_samples,
sample_negative=False):
# identify all cuts ("stitches") with a dictionary mapping exit wires to init wires
stitches = _identify_stitches(wire_path_map, circuit_wires)
# determine metadata about the distributions
frag_metadata = _get_distribution_metadata(frag_dists)
_, dist_obj_type, dist_dat_type = frag_metadata
# identify permutation that we will have to apply to the sampled states
qubit_perm = _united_axis_permutation(wire_path_map, circuit_wires,
frag_wires)
# figure out how to get various info from distributions
if dist_obj_type is tf.SparseTensor:
def _norm(dist): # normalization of a quasi-probability distribution
return sum(dist.values)
def _indices(dist): # number of indices in the distribution
return len(dist.indices)
def _probs(dist): # normalized 1-D array of probabilities
return dist.values.numpy() / sum(dist.values)
def _state(dist,
idx): # the state at a particular index for the 1-D array
return tuple(dist.indices.numpy()[idx])
else:
def _norm(dist):
return dist.numpy().sum()
def _indices(dist):
return np.prod(dist.shape)
def _probs(dist):
return dist.numpy().flatten() / _norm(dist)
def _state(dist, idx):
return tuple(int(bb) for bb in format(idx, f"0{len(dist.shape)}b"))
# pick one sample from a distribution
def _sample_dist(dist):
idx = np.random.choice(_indices(dist), p=_probs(dist))
return _state(dist, idx)
# get SIC-basis distributions
if _is_complex(dist_dat_type):
frag_dists_SIC = [
frag_dist.to_probabilities(SIC, SIC) for frag_dist in frag_dists
]
else:
frag_dists_SIC = [
frag_dist.shuffle_bases(SIC, SIC) for frag_dist in frag_dists
]
# if necessary, build an empty united distribution
if not num_samples:
frag_metadata_SIC = _get_distribution_metadata(frag_dists_SIC)
full_dist, _, _ = _get_uniting_objects(frag_dists_SIC, stitches,
frag_metadata_SIC)
# determine assignments of SIC-I operators that yield positive terms
positive_op_assignments \
= [ ops for ops in set_product(basis_ops_SIC + ["I"], repeat = len(stitches))
if sum([ op == "I" for op in ops ]) % 2 == sample_negative ]
# compute norms for all assigments of SIC-basis operators at all cut locations
norms = {}
for op_assignment in positive_op_assignments:
dist_factors = _collect_tensor_factors(frag_dists_SIC, stitches,
op_assignment)
term_norm = np.prod([_norm(dist) for dist in dist_factors])
scalar_factor = np.prod(
[1 if op == "I" else 3 / 2 for op in op_assignment])
norms[op_assignment] = scalar_factor * term_norm
if not num_samples:
full_dist += scalar_factor * reduce(tf_outer_product,
dist_factors[::-1])
total_norm = sum(norms.values())
# return the united distribution if appropriate
if not num_samples:
return tf_transpose(full_dist, qubit_perm) / total_norm, total_norm
# determine the term to sample from for each sample
term_probs = np.array(list(norms.values())) / total_norm
sample_term_indices = np.random.choice(len(term_probs),
num_samples,
p=term_probs)
sample_assignments = [
positive_op_assignments[idx] for idx in sample_term_indices
]
# collect a histogram of samples
samples = {}
for op_assignment in sample_assignments:
dist_factors = _collect_tensor_factors(frag_dists_SIC, stitches,
op_assignment)
frag_sample = tuple(val for dist in dist_factors[::-1]
for val in _sample_dist(dist))
circuit_sample = tuple(frag_sample[pp] for pp in qubit_perm)
try:
samples[circuit_sample] += 1
except:
samples[circuit_sample] = 1
return samples, total_norm
def main():
config = parse_arguments()
from sqlalchemy import create_engine
from sqlalchemy import Column, Table, MetaData
from sqlalchemy import String, Integer, Float
from sqlalchemy.orm import sessionmaker, mapper
from sqlalchemy.ext.declarative import declarative_base
from time import sleep, time
from stat_reader import StatReader
from itertools import product as set_product
from os import getuid
engine = create_engine('sqlite:///auto_test_log.sqlite')
metadata = MetaData(bind=engine)
Session = sessionmaker(bind=engine)
session = Session()
Base = declarative_base()
if getuid() != 0:
print "Must be root!!"
return
experiment_info = Table(
'experiment_info', metadata,
Column('experiment_id', String(32), primary_key=True),
Column('start_time', String()), Column('end_time', String()),
Column('total_time', Float()), Column('scheduler', String()),
Column('uname_a', String()), Column('cpufreq_governor', String()),
Column('mt_mc_state', String()), Column('target_load_level', String()),
Column('num_of_process', String()))
for c_name in proc_sys_kernel_olord_files:
experiment_info.append_column(Column(c_name, Float))
for c_name in sys_ondemand_files:
experiment_info.append_column(Column(c_name, Float))
experiment_data = Table('experiment_data', metadata,
Column('data_id', String(48), primary_key=True),
Column('experiment_id', String(32)),
Column('time', Float()), Column('voltage', Float),
Column('current', Float), Column('power', Float),
Column('temperature', Float),
Column('cpus_online', String(10)))
metadata.create_all()
class ExperimentInfo(Base):
__table__ = experiment_info
class ExperimentData(Base):
__table__ = experiment_data
eii = experiment_info.insert()
eid = experiment_data.insert()
psl = PowerSerialLogger()
tl = TemperatureLogger()
ocl = OnlineCPULogger()
def perform_experiment_a(experiment_config):
from md5 import new as new_md5
from pickle import dumps as pickle_dump
from platform import uname
from datetime import datetime
from threading import Thread
from pprint import PrettyPrinter
from simple_run import run_test
pp = PrettyPrinter(indent=4)
print "Running experiment with config:"
pp.pprint(experiment_config)
current_experiment_id = new_md5(
pickle_dump(experiment_config) + str(datetime.now())).hexdigest()
info_dict = {
'experiment_id': current_experiment_id,
'start_time': str(datetime.now()),
'scheduler': experiment_config['scheduler'],
'uname_a': ' '.join(uname()),
'cpufreq_governor': get_governor_string(),
'mt_mc_state': get_smt_mc_power_savings(),
'target_load_level': experiment_config['target_load_level'],
'num_of_process': experiment_config['num_of_process'],
'num_of_ops': experiment_config['num_of_ops']
}
info_dict.update(get_kernel_olord_settings())
info_dict.update(get_ondemand_settings())
eii.execute(info_dict)
test_args = {
'n': experiment_config['num_of_process'],
'a': 'dont_set',
'l': experiment_config['target_load_level'],
'o': experiment_config['num_of_ops']
}
class TestThread(Thread):
def __init__(self, args):
self.args = args
Thread.__init__(self)
def run(self):
self.total_time = run_test(self.args)
tt = TestThread(test_args)
tt.start()
start_time = time()
data_counter = 0
while tt.is_alive():
t = time() - start_time
data_dict = {
'data_id': current_experiment_id + str(data_counter).zfill(16),
'experiment_id': current_experiment_id,
'time': t,
'voltage': psl.voltage,
'current': psl.current,
'power': psl.power,
'temperature': tl.temperature,
'cpus_online': ocl.online_cpus
}
eid.execute(data_dict)
sleep_time = 1.0 - time() + (t + start_time)
sleep_time = 0.0 if (sleep_time < 0) else sleep_time
#print sleep_time
sleep(sleep_time)
data_counter += 1
r = session.query(ExperimentInfo).filter_by(
experiment_id=current_experiment_id).first()
r.end_time = str(datetime.now())
r.total_time = float(tt.total_time)
session.commit()
iter_set = set_product([80], [10], [10], [30], [80], range(90, 9, -5),
range(10))
iter_set = list(iter_set)
test_size = len(iter_set)
start_time = time()
counter = 0
for i in iter_set:
sleep(10.0)
if psl.power < 0.1:
print "Power ridiculously small, check connection to power source."
return
set_config_file(
'/sys/devices/system/cpu/cpufreq/ondemand/hotplug_in_load_limit',
i[0])
set_config_file(
'/sys/devices/system/cpu/cpufreq/ondemand/hotplug_out_load_limit',
i[1])
set_config_file(
'/sys/devices/system/cpu/cpufreq/ondemand/hotplug_in_sampling_period',
i[2])
set_config_file(
'/sys/devices/system/cpu/cpufreq/ondemand/hotplug_out_sampling_period',
i[2])
set_config_file(
'/sys/devices/system/cpu/cpufreq/ondemand/up_threshold', i[3])
set_config_file('/proc/sys/kernel/sched_olord_lb_upper_limit', i[4])
counter += 1
print "Starting test #%i of ~%i" % (counter, test_size)
t = time() - start_time
r = float(counter) / float(test_size)
eta = t / r - t
print "Estimated time left:", eta, "s"
perform_experiment_a({
'scheduler': 'CFS',
'target_load_level': i[5] / 100.0,
'num_of_process': 4,
'num_of_ops': 100000
})
def _label(spokes, lattice_size):
return f"$N={lattice_size}$"
elif len(spoke_vals) > 1 and len(lattice_size_vals) == 1:
title_text = f"$N={lattice_size_vals[0]}$"
def _label(spokes, lattice_size):
return f"$q={spokes}$"
else:
title_text = f"$q={spoke_vals[0]}$, $N={lattice_size_vals[0]}$"
def _label(*args):
return None
make_legend = False
for spokes, lattice_size in set_product(spoke_vals, lattice_size_vals):
volume = lattice_size**dimensions
label = _label(spokes, lattice_size)
dat_base_dir = os.path.join(root_dir, dat_dir)
dat_file_name = dat_name_builder(dat_base_dir, spokes, lattice_size,
dimensions, network_type, test_run)
try:
log_probs = np.loadtxt(dat_file_name("log_probs"))
log_norms = np.loadtxt(dat_file_name("log_norms"))
except OSError:
print("data files not found:")
print(dat_file_name("log_probs"))
print(dat_file_name("log_norms"))
def _checkerboard_lattice_vectors(lattice_shape):
assert (all([num % 2 == 0
for num in lattice_shape])) # even-sized lattices only
signed_ones = set_product([+1, -1], repeat=len(lattice_shape) - 1)
return [(1, ) + vec for vec in signed_ones]
def sample_separable_distribution(frag_amps, wire_path_map, circuit_wires,
frag_wires, num_samples):
# identify all cuts ("stitches") with a dictionary mapping exit wires to init wires
stitches = _identify_stitches(wire_path_map, circuit_wires)
# determine metadata about the distributions
frag_metadata = _get_distribution_metadata(frag_amps)
united_dist_shape, dist_obj_type, dist_dat_type = frag_metadata
assert (_is_complex(dist_dat_type))
assert (not dist_obj_type is tf.SparseTensor)
# identify permutation that we will have to apply to the sampled states
qubit_perm = _united_axis_permutation(wire_path_map, circuit_wires,
frag_wires)
# figure out how to get various info from amplitude vectors
def _norm(amps):
return (abs(amps.numpy())**2).sum()
def _indices(amps):
return np.prod(amps.shape)
def _probs(amps):
return abs(amps.numpy().flatten())**2 / _norm(amps)
def _state(bits, idx):
return tuple(int(bb) for bb in format(idx, f"0{bits}b"))
# pick one sample from a distribution
def _sample_dist(dist):
idx = np.random.choice(_indices(dist), p=_probs(dist))
return _state(len(dist.shape), idx)
# if necessary, build an empty united distribution
if not num_samples:
real_zero = abs(tf.zeros(1, dtype=dist_dat_type))
full_dist = tf.zeros(united_dist_shape, dtype=real_zero.dtype)
# compute norms for all assigments of operators at all cut locations
norms = {}
for op_assignment in set_product([0, 1], repeat=len(stitches)):
dist_factors = _collect_tensor_factors(frag_amps, stitches,
op_assignment)
norms[op_assignment] = np.prod([_norm(dist) for dist in dist_factors])
if not num_samples:
full_dist += abs(reduce(tf_outer_product, dist_factors[::-1]))**2
# return the united distribution if appropriate
if not num_samples:
return tf_transpose(full_dist, qubit_perm)
# determine the term to sample from for each sample
term_probs = np.array(list(norms.values()))
sample_term_indices = np.random.choice(len(term_probs),
num_samples,
p=term_probs)
sample_assignments = [
_state(len(stitches), idx) for idx in sample_term_indices
]
# collect a histogram of samples
samples = {}
for op_assignment in sample_assignments:
dist_factors = _collect_tensor_factors(frag_amps, stitches,
op_assignment)
frag_sample = tuple(val for dist in dist_factors[::-1]
for val in _sample_dist(dist))
circuit_sample = tuple(frag_sample[pp] for pp in qubit_perm)
try:
samples[circuit_sample] += 1
except:
samples[circuit_sample] = 1
return samples
def _get_single_fragment_probabilities(fragment, init_wires = None, exit_wires = None,
backend_simulator = "statevector_simulator",
init_basis = pauli, exit_basis = pauli,
dtype = tf.float64, **kwargs):
# save the bases in which we want to store init/exit conditions
final_init_basis = init_basis
final_exit_basis = exit_basis
# we can only actually *compute* distributions in certain bases,
# so if we weren't asked for one of those, choose one of them to use for now
if init_basis not in [ SIC, IZXY ]:
init_basis = SIC
if exit_basis != IZXY:
exit_basis = IZXY
'''
this method accepts:
(i) a circuit fragment,
(ii) a list of "input" wires (`init_wires`),
(iii) a list of "output" wires (`exit_wires`),
(iv) additional keyword arguments passed to the Qiskit statevector simulator.
this method returns probability distributions for non-exit-wires of the fragment
'''
if backend_simulator == "statevector_simulator":
frag_amps = _get_single_fragment_amplitudes(fragment, init_wires, exit_wires)
return frag_amps.to_probabilities(dtype = dtype)
def _dist_terms_IZXY(state):
if state[0] == "-": return [ "I" ]
else: return [ state, "I" ]
# initialize an empty conditional distribution over measurement outcomes
frag_dist = FragmentProbabilities(init_basis = init_basis)
# identify the axes we will project out for the exit wires
# and sort exit wires according to these axes
exit_axes = { wire : -fragment.qubits.index(wire)-1 for wire in exit_wires }
exit_wires = sorted(exit_wires, key = lambda wire : exit_axes[wire] )
# sort init wires to fix their order
init_wires = sorted(init_wires, key = lambda wire : fragment.qubits.index(wire))
if init_basis == SIC:
init_state_basis = list(state_vecs_SIC.keys())
else: # init_basis == IZXY
init_state_basis = list(state_vecs_ZXY.keys())
# remember the number of shots we were told to run
shots = kwargs["shots"]
del kwargs["shots"]
# for every choice of states prepared on all on init wires
for init_states in set_product(init_state_basis, repeat = len(init_wires)):
init_shots = shots # number of shots for this set of init states
if init_basis == SIC:
init_dist = frag_dist
else: # init_basis == IZXY
init_dist = FragmentProbabilities(init_basis)
# if we are simulating with initial states polarized in - Z/X/Y
# then we are actually collecting data for an insertion of I,
# so we only need a third of the number of shots (per such state)
init_shots /= 3**sum( state[0] == "-" for state in init_states )
if init_shots < 1 : continue
init_shots = int(init_shots + 0.5)
# build circuit with the "input" states prepared appropriately
init_conds = frozenset(zip(init_wires, init_states))
prep_circuits = [ _act_gate(fragment, _prep_gate(state), wire)
for wire, state in init_conds ]
init_circuit = reduce(lambda x, y : x + y, prep_circuits + [ fragment ])
# for every choice of measurement bases on all on exit wires
for exit_bases in set_product(_basis_gates.keys(), repeat = len(exit_wires)):
# build a circuit to measure in the correct bases
measurement_circuit = [ _act_gate(fragment, _basis_gates[basis], wire)
for wire, basis in zip(exit_wires, exit_bases) ]
circuit = reduce(lambda x, y : x + y,
[ init_circuit ] + measurement_circuit)
# get probability distribution over measurement outcomes
full_dist = get_circuit_probabilities(circuit, backend_simulator,
dtype = dtype, shots = init_shots,
**kwargs)
# project onto given exit-wire measurement outcomes
for exit_bits in set_product(range(2), repeat = len(exit_wires)):
exit_states = [ ( "+" if bit == 0 else "-" ) + basis
for bit, basis in zip(exit_bits, exit_bases) ]
projected_dist = deepcopy(full_dist)
for wire, bit_state in zip(exit_wires, exit_bits):
qubits = len(projected_dist.shape)
begin = [ 0 ] * qubits
size = [ 2 ] * qubits
begin[exit_axes[wire]] = bit_state
size[exit_axes[wire]] = 1
projected_dist = tf.sparse.slice(projected_dist, begin, size)
projected_dist = tf.sparse.reshape(projected_dist, (2,)*(qubits-1))
# loop over all assignments of + Z/X/Y and I at exit wires
exit_terms = [ _dist_terms_IZXY(state) for state in exit_states ]
for exit_states in set_product(*exit_terms):
# divide distribution by 3 for each identity operator (I)
# to average over measurements in 3 different bases
iden_fac = 3**sum( state == "I" for state in exit_states )
exit_conds = zip(exit_wires, exit_states)
init_dist.add(init_conds, exit_conds, projected_dist / iden_fac)
### end construction of init_dist
if init_dist is frag_dist: continue
for conditions, dist in init_dist:
init_conds = set( cond for cond in conditions if cond[0] )
exit_conds = set( cond for cond in conditions if not cond[0] )
if not init_conds:
frag_dist.add(conditions, dist)
continue
init_terms = [ _dist_terms_IZXY(state) for state in init_states ]
for init_ops in set_product(*init_terms):
iden_fac = 3**sum( op == "I" for op in init_ops )
init_conds = zip(init_wires, init_ops)
frag_dist.add(init_conds, exit_conds, dist / iden_fac)
# if we computed distributions in the same init/exit bases as we were asked for,
# then just return the conditional distribution we computed
# otherwise, change bases appropriately
if init_basis == final_init_basis and exit_basis == final_exit_basis:
return frag_dist
else:
return frag_dist.shuffle_bases(final_init_basis, final_exit_basis)
def main():
config = parse_arguments()
from sqlalchemy import create_engine
from sqlalchemy import Column, Table, MetaData
from sqlalchemy import String, Integer, Float
from sqlalchemy.orm import sessionmaker, mapper
from sqlalchemy.ext.declarative import declarative_base
from time import sleep, time
from stat_reader import StatReader
from itertools import product as set_product
from os import getuid
engine = create_engine('sqlite:///auto_test_log.sqlite')
metadata = MetaData(bind=engine)
Session = sessionmaker(bind=engine)
session = Session()
Base = declarative_base()
if getuid() != 0:
print "Must be root!!"
return
experiment_info = Table('experiment_info', metadata,
Column('experiment_id', String(32), primary_key=True),
Column('start_time', String()),
Column('end_time', String()),
Column('total_time', Float()),
Column('scheduler', String()),
Column('uname_a', String()),
Column('cpufreq_governor', String()),
Column('mt_mc_state', String()),
Column('target_load_level', String()),
Column('num_of_process', String()))
for c_name in proc_sys_kernel_olord_files:
experiment_info.append_column(Column(c_name, Float))
for c_name in sys_ondemand_files:
experiment_info.append_column(Column(c_name, Float))
experiment_data = Table('experiment_data', metadata,
Column('data_id', String(48), primary_key=True),
Column('experiment_id', String(32)),
Column('time', Float()),
Column('voltage', Float),
Column('current', Float),
Column('power', Float),
Column('temperature', Float),
Column('cpus_online', String(10)))
metadata.create_all()
class ExperimentInfo(Base):
__table__ = experiment_info
class ExperimentData(Base):
__table__ = experiment_data
eii = experiment_info.insert()
eid = experiment_data.insert()
psl = PowerSerialLogger()
tl = TemperatureLogger()
ocl = OnlineCPULogger()
def perform_experiment_a(experiment_config):
from md5 import new as new_md5
from pickle import dumps as pickle_dump
from platform import uname
from datetime import datetime
from threading import Thread
from pprint import PrettyPrinter
from simple_run import run_test
pp = PrettyPrinter(indent = 4)
print "Running experiment with config:"
pp.pprint(experiment_config)
current_experiment_id = new_md5(pickle_dump(experiment_config) + str(datetime.now())).hexdigest()
info_dict = {'experiment_id': current_experiment_id,
'start_time': str(datetime.now()),
'scheduler': experiment_config['scheduler'],
'uname_a': ' '.join(uname()),
'cpufreq_governor': get_governor_string(),
'mt_mc_state': get_smt_mc_power_savings(),
'target_load_level': experiment_config['target_load_level'],
'num_of_process': experiment_config['num_of_process'],
'num_of_ops': experiment_config['num_of_ops']}
info_dict.update(get_kernel_olord_settings())
info_dict.update(get_ondemand_settings())
eii.execute(info_dict)
test_args = {'n': experiment_config['num_of_process'],
'a': 'dont_set',
'l': experiment_config['target_load_level'],
'o': experiment_config['num_of_ops']}
class TestThread(Thread):
def __init__(self, args):
self.args = args
Thread.__init__(self)
def run(self):
self.total_time = run_test(self.args)
tt = TestThread(test_args)
tt.start()
start_time = time()
data_counter = 0
while tt.is_alive():
t = time() - start_time
data_dict = {'data_id': current_experiment_id + str(data_counter).zfill(16),
'experiment_id': current_experiment_id,
'time': t,
'voltage': psl.voltage,
'current': psl.current,
'power': psl.power,
'temperature': tl.temperature,
'cpus_online': ocl.online_cpus}
eid.execute(data_dict)
sleep_time = 1.0 - time() + (t + start_time)
sleep_time = 0.0 if (sleep_time < 0) else sleep_time
#print sleep_time
sleep(sleep_time)
data_counter += 1
r = session.query(ExperimentInfo).filter_by(experiment_id = current_experiment_id).first()
r.end_time = str(datetime.now())
r.total_time = float(tt.total_time)
session.commit()
iter_set = set_product([80],
[10],
[10],
[30],
[80],
range(90,9,-5),
range(10))
iter_set = list(iter_set)
test_size = len(iter_set)
start_time = time()
counter = 0
for i in iter_set:
sleep(10.0)
if psl.power < 0.1:
print "Power ridiculously small, check connection to power source."
return
set_config_file('/sys/devices/system/cpu/cpufreq/ondemand/hotplug_in_load_limit', i[0])
set_config_file('/sys/devices/system/cpu/cpufreq/ondemand/hotplug_out_load_limit', i[1])
set_config_file('/sys/devices/system/cpu/cpufreq/ondemand/hotplug_in_sampling_period', i[2])
set_config_file('/sys/devices/system/cpu/cpufreq/ondemand/hotplug_out_sampling_period', i[2])
set_config_file('/sys/devices/system/cpu/cpufreq/ondemand/up_threshold', i[3])
set_config_file('/proc/sys/kernel/sched_olord_lb_upper_limit', i[4])
counter += 1
print "Starting test #%i of ~%i" % (counter, test_size)
t = time() - start_time
r = float(counter) / float(test_size)
eta = t/r - t
print "Estimated time left:", eta, "s"
perform_experiment_a({'scheduler': 'CFS',
'target_load_level': i[5]/100.0,
'num_of_process': 4,
'num_of_ops': 100000})