def merge(i_left, i_right, i_end):
i0 = MemValue(i_left)
i1 = MemValue(i_right)
@for_range(i_left, i_end)
def loop(j):
if_then(and_(lambda: i0 < i_right,
or_(lambda: i1 >= i_end,
lambda: regint(reveal(A[i0] <= A[i1])))))
B[j] = A[i0]
i0.iadd(1)
else_then()
B[j] = A[i1]
i1.iadd(1)
end_if()
def twos_complement(x):
bits = x.bit_decompose(k)[::-1]
bit_array = Array(k, cint)
bit_array.assign(bits)
twos_result = MemValue(cint(0))
@for_range(k)
def block(i):
val = twos_result.read()
val <<= 1
val += 1 - bit_array[i]
twos_result.write(val)
return twos_result.read() + 1
def twos_complement(x):
bits = x.bit_decompose(k)[::-1]
bit_array = Array(k, cint)
bit_array.assign(bits)
twos_result = MemValue(cint(0))
@for_range(k)
def block(i):
val = twos_result.read()
val <<= 1
val += 1 - bit_array[i]
twos_result.write(val)
return twos_result.read() + 1
def mergesort(A):
B = Array(len(A), sint)
def merge(i_left, i_right, i_end):
i0 = MemValue(i_left)
i1 = MemValue(i_right)
@for_range(i_left, i_end)
def loop(j):
if_then(and_(lambda: i0 < i_right,
or_(lambda: i1 >= i_end,
lambda: regint(reveal(A[i0] <= A[i1])))))
B[j] = A[i0]
i0.iadd(1)
else_then()
B[j] = A[i1]
i1.iadd(1)
end_if()
width = MemValue(1)
@do_while
def width_loop():
@for_range(0, len(A), 2 * width)
def merge_loop(i):
merge(i, i + width, i + 2 * width)
A.assign(B)
width.imul(2)
return width < len(A)
def approximate_reciprocal(divisor, k, f, theta):
"""
returns aproximation of 1/divisor
where type(divisor) = cint
"""
def twos_complement(x):
bits = x.bit_decompose(k)[::-1]
bit_array = Array(k, cint)
bit_array.assign(bits)
twos_result = MemValue(cint(0))
@for_range(k)
def block(i):
val = twos_result.read()
val <<= 1
val += 1 - bit_array[i]
twos_result.write(val)
return twos_result.read() + 1
bit_array = Array(k, cint)
bits = divisor.bit_decompose(k)[::-1]
bit_array.assign(bits)
cnt_leading_zeros = MemValue(regint(0))
flag = MemValue(regint(0))
cnt_leading_zeros = MemValue(regint(0))
normalized_divisor = MemValue(divisor)
@for_range(k)
def block(i):
flag.write(flag.read() | bit_array[i] == 1)
@if_(flag.read() == 0)
def block():
cnt_leading_zeros.write(cnt_leading_zeros.read() + 1)
normalized_divisor.write(normalized_divisor << 1)
q = MemValue(two_power(k))
e = MemValue(twos_complement(normalized_divisor.read()))
@for_range(theta)
def block(i):
qread = q.read()
eread = e.read()
qread += (qread * eread) >> k
eread = (eread * eread) >> k
q.write(qread)
e.write(eread)
res = q >> (2 * k - 2 * f - cnt_leading_zeros)
return res
def map_reduce_single(n_parallel, n_loops, initializer, reducer, mem_state=None):
if not isinstance(n_parallel, int):
raise CompilerException('Number of parallel executions' \
'must be constant')
n_parallel = n_parallel or 1
if mem_state is None:
# default to list of MemValues to allow varying types
mem_state = [MemValue(x) for x in initializer()]
use_array = False
else:
# use Arrays for multithread version
use_array = True
def decorator(loop_body):
if isinstance(n_loops, int):
loop_rounds = n_loops / n_parallel \
if n_parallel < n_loops else 0
else:
loop_rounds = n_loops / n_parallel
def write_state_to_memory(r):
if use_array:
mem_state.assign(r)
else:
# cannot do mem_state = [...] due to scope issue
for j,x in enumerate(r):
mem_state[j].write(x)
# will be optimized out if n_loops <= n_parallel
@for_range(loop_rounds)
def f(i):
state = tuplify(initializer())
for k in range(n_parallel):
j = i * n_parallel + k
state = reducer(tuplify(loop_body(j)), state)
r = reducer(mem_state, state)
write_state_to_memory(r)
if isinstance(n_loops, int):
state = mem_state
for j in range(loop_rounds * n_parallel, n_loops):
state = reducer(tuplify(loop_body(j)), state)
else:
@for_range(loop_rounds * n_parallel, n_loops)
def f(j):
r = reducer(tuplify(loop_body(j)), mem_state)
write_state_to_memory(r)
state = mem_state
for i,x in enumerate(state):
if use_array:
mem_state[i] = x
else:
mem_state[i].write(x)
def returner():
return untuplify(tuple(state))
return returner
return decorator
def approximate_reciprocal(divisor, k, f, theta):
"""
returns aproximation of 1/divisor
where type(divisor) = cint
"""
def twos_complement(x):
bits = x.bit_decompose(k)[::-1]
bit_array = Array(k, cint)
bit_array.assign(bits)
twos_result = MemValue(cint(0))
@for_range(k)
def block(i):
val = twos_result.read()
val <<= 1
val += 1 - bit_array[i]
twos_result.write(val)
return twos_result.read() + 1
bit_array = Array(k, cint)
bits = divisor.bit_decompose(k)[::-1]
bit_array.assign(bits)
cnt_leading_zeros = MemValue(regint(0))
flag = MemValue(regint(0))
cnt_leading_zeros = MemValue(regint(0))
normalized_divisor = MemValue(divisor)
@for_range(k)
def block(i):
flag.write(flag.read() | bit_array[i] == 1)
@if_(flag.read() == 0)
def block():
cnt_leading_zeros.write(cnt_leading_zeros.read() + 1)
normalized_divisor.write(normalized_divisor << 1)
q = MemValue(two_power(k))
e = MemValue(twos_complement(normalized_divisor.read()))
qr = q.read()
er = e.read()
for i in range(theta):
qr = qr + shift_two(qr * er, k)
er = shift_two(er * er, k)
q = qr
res = shift_two(q, (2*k - 2*f - cnt_leading_zeros))
return res
def test_while():
num_vals = 5
counter = MemValue(sint(num_vals - 1))
source_arr = Array(num_vals, sint)
for i in range(num_vals):
source_arr[i] = sint(i)
target_arr = Array(num_vals, sint)
@do_while
def body():
counter_val = counter.read()
counter_val_open = counter_val.reveal()
target_arr[counter_val_open] = source_arr[counter_val_open] + 1
counter.write(counter_val - 1)
opened = counter.reveal()
return opened >= 0
runtime_assert_arr_equals([1, 2, 3, 4, 5], target_arr,
default_test_name())
class bits(Tape.Register):
n = 40
size = 1
PreOp = staticmethod(floatingpoint.PreOpN)
MemValue = staticmethod(lambda value: MemValue(value))
decomposed = None
@staticmethod
def PreOR(l):
return [1 - x for x in \
floatingpoint.PreOpN(operator.mul, \
[1 - x for x in l])]
@classmethod
def get_type(cls, length):
if length is None:
return cls
elif length == 1:
return cls.bit_type
if length not in cls.types:
class bitsn(cls):
n = length
cls.types[length] = bitsn
bitsn.__name__ = cls.__name__ + str(length)
return cls.types[length]
@classmethod
def conv(cls, other):
if isinstance(other, cls):
return other
elif isinstance(other, MemValue):
return cls.conv(other.read())
else:
res = cls()
res.load_other(other)
return res
hard_conv = conv
@classmethod
def compose(cls, items, bit_length=1):
return cls.bit_compose(
sum([util.bit_decompose(item, bit_length) for item in items], []))
@classmethod
def bit_compose(cls, bits):
if len(bits) == 1:
return bits[0]
bits = list(bits)
res = cls.new(n=len(bits))
cls.bitcom(res, *(sbit.conv(bit) for bit in bits))
res.decomposed = bits
return res
def bit_decompose(self, bit_length=None):
n = bit_length or self.n
suffix = [0] * (n - self.n)
if n == 1 and self.n == 1:
return [self]
n = min(n, self.n)
if self.decomposed is None or len(self.decomposed) < n:
res = [self.bit_type() for i in range(n)]
self.bitdec(self, *res)
self.decomposed = res
return res + suffix
else:
return self.decomposed[:n] + suffix
@classmethod
def load_mem(cls, address, mem_type=None):
res = cls()
if mem_type == 'sd':
return cls.load_dynamic_mem(address)
else:
cls.load_inst[util.is_constant(address)](res, address)
return res
def store_in_mem(self, address):
self.store_inst[isinstance(address, (int, long))](self, address)
def __init__(self, value=None, n=None, size=None):
if size != 1 and size is not None:
raise Exception('invalid size for bit type: %s' % size)
Tape.Register.__init__(self, self.reg_type, Program.prog.curr_tape)
self.set_length(n or self.n)
if value is not None:
self.load_other(value)
def set_length(self, n):
if n > self.max_length:
print self.max_length
raise Exception('too long: %d' % n)
self.n = n
def load_other(self, other):
if isinstance(other, (int, long)):
self.set_length(self.n or util.int_len(other))
self.load_int(other)
elif isinstance(other, regint):
self.conv_regint(self.n, self, other)
elif isinstance(self, type(other)) or isinstance(other, type(self)):
self.mov(self, other)
else:
try:
other = self.bit_compose(other.bit_decompose())
self.mov(self, other)
except:
raise CompilerError('cannot convert from %s to %s' % \
(type(other), type(self)))
def long_one(self):
return 2**self.n - 1
def __repr__(self):
return '%s(%d/%d)' % \
(super(bits, self).__repr__(), self.n, type(self).n)
def memorize(x):
if isinstance(x, (tuple, list)):
return tuple(memorize(i) for i in x)
else:
return MemValue(x)
def int2FL_plain(a, gamma, l, kappa):
lam = gamma - 1
a_abs = 0
v = cint(0)
p = cint(0)
s = cint(0)
z = cint(0)
# extracts the sign and calculates the abs
s = cint(a < 0)
a_abs = a * (1 - 2 * s)
# isolates most significative bit
a_bits = a_abs.bit_decompose()
b = 0
b_c = 1
blen = 0
for a_i in range(len(a_bits) - 1, -1, -1): # enumerate(a_bits):
b = (a_bits[a_i]) * (b == 0) * ((b_c) / 2) + b
blen = (a_bits[a_i]) * (blen == 0) * ((a_i + 1)) + blen
b_c = b_c * 2
# obtains p
# blen= len(a_bits) - blen
v = a_abs * b # (2 ** (b))#scale a
p = - (lam - blen) # (len(a_bits)-blen))
# reduces v
v_l = MemValue(v)
z_l = MemValue(z)
if_then(a_abs > 0)
if (lam > l):
v_l.write(v_l.read() / (2 ** (gamma - l - 1)))
else:
v_l.write(v_l.read() * (2 ** l - lam))
else_then()
z_l.write(cint(1))
end_if()
# corrects output
# s is coming from the abs extraction
v = cint(v_l.read())
z = cint(z_l.read())
p = cint((p + lam - l) * (1 - z))
return v, p, z, s
def decorator(loop_body):
my_n_parallel = n_parallel
if isinstance(n_parallel, int):
if isinstance(n_loops, int):
loop_rounds = n_loops / n_parallel \
if n_parallel < n_loops else 0
else:
loop_rounds = n_loops / n_parallel
def write_state_to_memory(r):
if use_array:
mem_state.assign(r)
else:
# cannot do mem_state = [...] due to scope issue
for j, x in enumerate(r):
mem_state[j].write(x)
if n_parallel is not None:
# will be optimized out if n_loops <= n_parallel
@for_range(loop_rounds)
def f(i):
state = tuplify(initializer())
for k in range(n_parallel):
j = i * n_parallel + k
state = reducer(tuplify(loop_body(j)), state)
r = reducer(mem_state, state)
write_state_to_memory(r)
else:
n_parallel_reg = MemValue(regint(0))
parent_block = get_block()
@while_do(lambda x: x + n_parallel_reg <= n_loops, regint(0))
def _(i):
state = tuplify(initializer())
k = 0
block = get_block()
while k < n_loops and (len(get_block()) < get_program().budget \
or k == 0) \
and block is get_block():
j = i + k
state = reducer(tuplify(loop_body(j)), state)
k += 1
r = reducer(mem_state, state)
write_state_to_memory(r)
global n_opt_loops
n_opt_loops = k
n_parallel_reg.write(k)
return i + k
my_n_parallel = n_opt_loops
loop_rounds = n_loops / my_n_parallel
blocks = get_tape().basicblocks
n_to_merge = 5
if loop_rounds == 1 and parent_block is blocks[-n_to_merge]:
# merge blocks started by if and do_while
def exit_elimination(block):
if block.exit_condition is not None:
for reg in block.exit_condition.get_used():
reg.can_eliminate = True
exit_elimination(parent_block)
merged = parent_block
merged.exit_condition = blocks[-1].exit_condition
merged.exit_block = blocks[-1].exit_block
assert parent_block is blocks[-n_to_merge]
assert blocks[-n_to_merge + 1] is \
get_tape().req_node.children[-1].nodes[0].blocks[0]
for block in blocks[-n_to_merge + 1:]:
merged.instructions += block.instructions
exit_elimination(block)
del blocks[-n_to_merge + 1:]
del get_tape().req_node.children[-1]
merged.children = []
get_tape().active_basicblock = merged
else:
req_node = get_tape().req_node.children[-1].nodes[0]
req_node.children[0].aggregator = lambda x: loop_rounds * x[0]
if isinstance(n_loops, int):
state = mem_state
for j in range(loop_rounds * my_n_parallel, n_loops):
state = reducer(tuplify(loop_body(j)), state)
else:
@for_range(loop_rounds * my_n_parallel, n_loops)
def f(j):
r = reducer(tuplify(loop_body(j)), mem_state)
write_state_to_memory(r)
state = mem_state
for i, x in enumerate(state):
if use_array:
mem_state[i] = x
else:
mem_state[i].write(x)
def returner():
return untuplify(tuple(state))
return returner
