def _pyske_bcast(input_list: PList):
size = input_list.distribution[0]
nprocs = len(par.procs())
return input_list\
.get_partition()\
.mapi(lambda pid, a_list: list(map(lambda _: a_list, par.procs())) if pid == 0 else []) \
.flatten(Distribution([nprocs if pid == 0 else 0 for pid in par.procs()])) \
.distribute(Distribution(map(lambda _: 1, par.procs())))\
.flatten(Distribution(map(lambda _: size, par.procs())))
def fft(input_list: PList[float]) -> PList[complex]:
# pylint: disable=unsubscriptable-object
"""
Return the Discrete Fourier Transform.
Examples::
>>> from pyske.core import PList
>>> fft(PList.init(lambda _: 1.0, 128)).to_seq()[0]
(128+0j)
:param input_list: a PySke list of floating point numbers
:return: a parallel list of complex numbers
"""
size = len(input_list)
log2_size = int(math.log2(size))
nprocs = len(par.procs())
log2_nprocs = int(math.log2(nprocs))
assert size == 2**log2_size
assert nprocs == 2**log2_nprocs
result = input_list.map(complex)
for index_j in range(0, log2_nprocs):
permutation = result.get_partition() \
.permute(partial(_bit_complement, log2_nprocs - index_j - 1)) \
.flatten()
result = permutation.map2i(partial(_combine, size, log2_size, index_j),
result)
for index_j in range(log2_nprocs, log2_size):
permutation = result.get_partition() \
.map(lambda l: l.permute(partial(_bit_complement, log2_size - index_j - 1))) \
.flatten()
result = permutation.map2i(partial(_combine, size, log2_size, index_j),
result)
return result
def test_gather_distr():
# pylint: disable=missing-docstring
data = generate_str_plist()
size = data.length()
dst = par.randpid()
res = get_distribution(data.gather(dst))
exp = [size if i == dst else 0 for i in par.procs()]
assert res == exp
def test_distribute_distr():
# pylint: disable=missing-docstring
data = generate_str_plist()
size = data.length()
dst = par.randpid()
exp = Distribution([0 for _ in par.procs()])
exp[dst] = size
res = get_distribution(data.distribute(exp))
assert res == exp
def test_balance_distr():
# pylint: disable=missing-docstring
data = generate_str_plist()
size = data.length()
dst = par.randpid()
distr = Distribution([0 for _ in par.procs()])
distr[dst] = size
res = get_distribution(data.distribute(distr).balance())
exp = Distribution.balanced(size)
assert res == exp
def test_distribute_data():
# pylint: disable=missing-docstring
dst = par.randpid()
data = generate_int_plist()
size = data.length()
distr = Distribution([0 for _ in par.procs()])
distr[dst] = size
res = data.distribute(distr).to_seq()
exp = SList(range(0, size))
assert res == exp
def bcast(input_list: PList, src_pid: int) -> PList:
"""
Broadcast the data at source processor to all processors.
Example::
>>> from pyske.core import PList, par
>>> bcast(PList.from_seq([42]), 0).to_seq() == \
list(map(lambda _: 42, par.procs()))
True
:param input_list: a parallel list.
:param src_pid: the source processor identifier.
Pre-condition: ``src_pid in par.procs()``
:return: a parallel list.
"""
assert src_pid in par.procs()
size = input_list.distribution[src_pid]
nprocs = len(par.procs())
return input_list \
.get_partition() \
.mapi(lambda pid, lst: list(map(lambda _: lst, par.procs())) if pid == src_pid else []) \
.flatten(Distribution([nprocs if pid == src_pid else 0 for pid in par.procs()])) \
.distribute(Distribution(map(lambda _: 1, par.procs()))) \
.flatten(Distribution(map(lambda _: size, par.procs())))
def _main():
size, num_iter, _ = util.standard_parse_command_line(data_arg=False)
assert _is_power_of_2(size), "The size should be a power of 2."
assert _is_power_of_2(len(
par.procs())), "The number of processors should be a power of 2."
input_list = PList.init(lambda _: 1.0, size)
timing = Timing()
gc.disable()
for iteration in range(1, 1 + num_iter):
timing.start()
result = fft(input_list)
timing.stop()
gc.collect()
result = result.to_seq()[0]
util.print_experiment(result, timing.get(), par.at_root, iteration)
def pssr(input_list: PList) -> PList:
"""
Sort the input list.
Sorts using ``<`` only.
Example::
>>> from pyske.core import PList
>>> pssr(PList.init(lambda i: 10-i, 10)).to_seq()
[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
:param input_list: a parallel list.
Pre-condition: the size of the local lists should be at least
equal to the number of processors.
:return: a sorted list that is a permutation of ``input_list``.
"""
nprocs = len(par.procs())
if nprocs == 1:
return input_list.get_partition().map(sorted).flatten()
for local_size in input_list.distribution:
assert local_size >= nprocs
def permutation(index: int):
return int(index / nprocs) + nprocs * (index % nprocs)
def _sample(list_to_sample):
if list_to_sample:
size = len(list_to_sample)
step = int(size / nprocs)
return list_to_sample[step:size:step]
return []
locally_sorted = input_list.get_partition().map(sorted)
first_samples = locally_sorted.map(_sample).gather(0).get_partition()
second_samples = bcast(first_samples.map(_merge).map(_sample), 0)
slices = locally_sorted.map2(_slice, second_samples).flatten()
result = slices.permute(permutation).get_partition().map(_merge).flatten()
return result
PARSER.add_argument("--test", help="choice of the test", type=int, default=2)
PARSER.add_argument("-v", help="verbose mode", action='store_true')
ARGS = PARSER.parse_args()
ITERATIONS = ARGS.iter
SIZE = ARGS.size
SEQ = ARGS.seq
TST = ARGS.test
VRB = ARGS.v
if VRB:
par.at_root(lambda:
print("Iterations:\t", ITERATIONS,
"\nSize:\t", SIZE,
"\nSeq: \t", SEQ,
"\nTest:\t", TST,
"\nNprocs:\t", len(par.procs())))
def _test_mmr_direct(lst):
lst1 = lst.map(_f_map)
lst2 = lst1.map(_f_map)
res = lst2.reduce(_f_reduce, 0)
return res
def _test_mr_direct(lst):
def fct(num):
return _f_map(_f_map(num))
lst1 = lst.map(fct)
res = lst1.reduce(_f_reduce, 0)
return res