def run_delta(delta_change, delta_type, change_set, type_set, search_config,
original_config, bitcode, original_score, inx, queue):
if len(delta_change) > 0:
# always reset to lowest precision
utilities.to_2nd_highest_precision(change_set, type_set)
# apply change for variables in delta
utilities.to_highest_precision(delta_change, delta_type)
# run config
result = utilities.run_config(search_config, original_config, bitcode, inx)
score = utilities.get_dynamic_score(inx)
if result == 1 and score < original_score:
queue.put([inx, score])
def run_delta(delta_change, delta_type, change_set, type_set, search_config,
original_config, bitcode, original_score, inx, queue):
if len(delta_change) > 0:
# always reset to lowest precision
utilities.to_2nd_highest_precision(change_set, type_set)
# apply change for variables in delta
utilities.to_highest_precision(delta_change, delta_type)
# run config
result = utilities.run_config(search_config, original_config, bitcode, inx)
score = utilities.get_dynamic_score(inx)
if result == 1 and score < original_score:
queue.put([inx, score])
def search_config_dis(change_set, type_set, search_config, original_config, bitcode, search_counter, inx, queue):
# search from bottom up
utilities.to_2nd_highest_precision(change_set, type_set)
if utilities.run_config(search_config, original_config, bitcode, search_counter) != 1:
search_counter = search_counter + 1
if len(change_set) > 1:
dd_search_config_dis(change_set, type_set, search_config, original_config, bitcode, 2, search_counter)
else:
utilities.to_highest_precision(change_set, type_set)
else:
search_counter = search_counter + 1
# remove types that cannot be changed
for i in xrange(0, len(change_set)):
if len(type_set[i]) > 0 and change_set[i]["type"] == type_set[i][-1]:
queue.put(inx + i)
def search_config(change_set, type_set, search_config, original_config, bitcode, original_score):
global search_counter
# search from bottom up
utilities.to_2nd_highest_precision(change_set, type_set)
if utilities.run_config(search_config, original_config, bitcode, search_counter) != 1 or utilities.get_dynamic_score(search_counter) > original_score:
search_counter = search_counter + 1
dd_search_config(change_set, type_set, search_config, original_config, bitcode, original_score, 2)
else:
search_counter = search_counter + 1
# remove types that cannot be changed
for i in xrange(0, len(change_set)):
if len(type_set[i]) > 0 and change_set[i]["type"] == type_set[i][-1]:
del(type_set[i][:])
# remove highest precision from each type vector
for type_vector in type_set:
if len(type_vector) > 0:
type_vector.pop()
def search_config(change_set, type_set, search_config, original_config, bitcode, original_score):
global search_counter
# search from bottom up
utilities.to_2nd_highest_precision(change_set, type_set)
if utilities.run_config(search_config, original_config, bitcode, search_counter) != 1 or utilities.get_dynamic_score(search_counter) > original_score:
search_counter = search_counter + 1
dd_search_config(change_set, type_set, search_config, original_config, bitcode, original_score, 2)
else:
search_counter = search_counter + 1
# remove types that cannot be changed
for i in xrange(0, len(change_set)):
if len(type_set[i]) > 0 and change_set[i]["type"] == type_set[i][-1]:
del(type_set[i][:])
# remove highest precision from each type vector
for type_vector in type_set:
if len(type_vector) > 0:
type_vector.pop()
def dd_search_config(change_set, type_set, search_config, original_config, bitcode, original_score, div):
global search_counter
global CPU_NO
#
# partition change_set into deltas and delta inverses
#
delta_change_set = []
delta_type_set = []
delta_inv_change_set = []
delta_inv_type_set = []
div_size = int(math.ceil(float(len(change_set))/float(div)))
for i in xrange(0, len(change_set), div_size):
delta_change = []
delta_type = []
delta_inv_change = []
delta_inv_type = []
for j in xrange(0, len(change_set)):
if j >= i and j < i+div_size:
delta_change.append(change_set[j])
delta_type.append(type_set[j])
else:
delta_inv_change.append(change_set[j])
delta_inv_type.append(type_set[j])
delta_change_set.append(delta_change)
delta_type_set.append(delta_type)
delta_inv_change_set.append(delta_inv_change)
delta_inv_type_set.append(delta_inv_type)
min_score = -1
#
# iterate through all delta and inverse delta set
# record delta set that passes
#
pass_inx = -1
inv_is_better = False
#
# iterate through all deltas and delta inverses
# run in parallel
#
start_inx = search_counter
queue = Queue()
inv_queue = Queue()
workers = []
for i in xrange(0, len(delta_change_set)):
workers.append(Process(target=run_delta, args=(delta_change_set[i],
delta_type_set[i], change_set, type_set, search_config, original_config, bitcode,
original_score, start_inx+i, queue)))
for i in xrange(0, len(delta_inv_change_set)):
workers.append(Process(target=run_delta, args=(delta_inv_change_set[i],
delta_inv_type_set[i], change_set, type_set, search_config, original_config, bitcode,
original_score, start_inx+len(delta_change_set)+i, inv_queue)))
# run in parallel maximum cpu_no processes at a time
cpu_no = CPU_NO
#multiprocessing.cpu_count()
# cpu_no = 1
loop = int(len(workers)/cpu_no)
for i in xrange(0, loop):
for j in xrange(i*cpu_no, min((i+1)*cpu_no, len(workers))):
workers[j].start()
for j in xrange(i*cpu_no, min((i+1)*cpu_no, len(workers))):
workers[j].join()
while not queue.empty():
result = queue.get()
if min_score == -1 or result[1] < min_score:
pass_inx = result[0] - start_inx
inv_is_better = False
while not inv_queue.empty():
result = inv_queue.get()
if min_score == -1 or result[1] < min_score:
pass_inx = result[0] - start_inx - len(delta_change_set)
inv_is_better = True
search_counter += len(delta_change_set) + len(delta_inv_change_set)
#
# recursively search in pass delta or pass delta inverse
# right now keep searching for the first pass delta or
# pass delta inverse; later on we will integrate cost
# model here
#
if pass_inx != -1:
pass_change_set = delta_inv_change_set[pass_inx] if inv_is_better else delta_change_set[pass_inx]
pass_type_set = delta_inv_type_set[pass_inx] if inv_is_better else delta_type_set[pass_inx]
if len(pass_change_set) > 1:
# always reset to lowest precision
utilities.to_2nd_highest_precision(change_set, type_set)
dd_search_config(pass_change_set, pass_type_set, search_config, original_config, bitcode, original_score, 2)
else:
utilities.to_2nd_highest_precision(change_set, type_set)
utilities.to_highest_precision(pass_change_set, pass_type_set)
return
#
# stop searching when division greater than change set size
#
if div >= len(change_set):
utilities.to_highest_precision(change_set, type_set)
return
else:
dd_search_config(change_set, type_set, search_config, original_config, bitcode, original_score, 2*div)
return
def dd_search_config_dis(change_set, type_set, search_config, original_config, bitcode, div, search_counter):
#
# partition change_set into deltas and delta inverses
#
delta_change_set = []
delta_type_set = []
delta_inv_change_set = []
delta_inv_type_set = []
div_size = int(math.ceil(float(len(change_set))/float(div)))
for i in xrange(0, len(change_set), div_size):
delta_change = []
delta_type = []
delta_inv_change = []
delta_inv_type = []
for j in xrange(0, len(change_set)):
if j >= i and j < i+div_size:
delta_change.append(change_set[j])
delta_type.append(type_set[j])
else:
delta_inv_change.append(change_set[j])
delta_inv_type.append(type_set[j])
delta_change_set.append(delta_change)
delta_type_set.append(delta_type)
delta_inv_change_set.append(delta_inv_change)
delta_inv_type_set.append(delta_inv_type)
#
# iterate through all delta and inverse delta set
# record delta set that passes
#
pass_inx = -1
inv_is_better = False
#
# iterate through all deltas and delta inverses
# run in parallel
#
for i in xrange(0, len(delta_change_set)):
delta_change = delta_change_set[i]
delta_type = delta_type_set[i]
if len(delta_change) > 0:
# always reset to lowest precision
utilities.to_2nd_highest_precision(change_set, type_set)
# apply change for variables in delta
utilities.to_highest_precision(delta_change, delta_type)
# run config
if utilities.run_config(search_config, original_config, bitcode, search_counter) == 1:
search_counter += 1
pass_inx = i
break
else:
search_counter += 1
if pass_inx == -1:
for i in xrange(0, len(delta_inv_change_set)):
delta_change = delta_inv_change_set[i]
delta_type = delta_inv_type_set[i]
if len(delta_change) > 0:
# always reset to lowest precision
utilities.to_2nd_highest_precision(change_set, type_set)
# apply change for variables in delta
utilities.to_highest_precision(delta_change, delta_type)
# run config
if utilities.run_config(search_config, original_config, bitcode, search_counter) == 1:
search_counter += 1
pass_inx = i
inv_is_better = True
break
else:
search_counter += 1
#
# recursively search in pass delta or pass delta inverse
# right now keep searching for the first pass delta or
# pass delta inverse; later on we will integrate cost
# model here
#
if pass_inx != -1:
pass_change_set = delta_inv_change_set[pass_inx] if inv_is_better else delta_change_set[pass_inx]
pass_type_set = delta_inv_type_set[pass_inx] if inv_is_better else delta_type_set[pass_inx]
if len(pass_change_set) > 1:
# always reset to lowest precision
utilities.to_2nd_highest_precision(change_set, type_set)
dd_search_config_dis(pass_change_set, pass_type_set, search_config, original_config, bitcode, 2, search_counter)
else:
utilities.to_2nd_highest_precision(change_set, type_set)
utilities.to_highest_precision(pass_change_set, pass_type_set)
return
#
# stop searching when division greater than change set size
#
if div >= len(change_set):
utilities.to_highest_precision(change_set, type_set)
return
else:
dd_search_config_dis(change_set, type_set, search_config, original_config, bitcode, 2*div, search_counter)
return
def dd_search_config(change_set, type_set, search_config, original_config, bitcode, original_score, div):
global search_counter
#
# partition change_set into deltas and delta inverses
#
delta_change_set = []
delta_type_set = []
delta_inv_change_set = []
delta_inv_type_set = []
div_size = int(math.ceil(float(len(change_set))/float(div)))
for i in xrange(0, len(change_set), div_size):
delta_change = []
delta_type = []
delta_inv_change = []
delta_inv_type = []
for j in xrange(0, len(change_set)):
if j >= i and j < i+div_size:
delta_change.append(change_set[j])
delta_type.append(type_set[j])
else:
delta_inv_change.append(change_set[j])
delta_inv_type.append(type_set[j])
delta_change_set.append(delta_change)
delta_type_set.append(delta_type)
delta_inv_change_set.append(delta_inv_change)
delta_inv_type_set.append(delta_inv_type)
min_score = -1
#
# iterate through all delta and inverse delta set
# record delta set that passes
#
pass_inx = -1
inv_is_better = False
#
# iterate through all deltas and delta inverses
# run in parallel
#
start_inx = search_counter
queue = Queue()
inv_queue = Queue()
workers = []
for i in xrange(0, len(delta_change_set)):
workers.append(Process(target=run_delta, args=(delta_change_set[i],
delta_type_set[i], change_set, type_set, search_config, original_config, bitcode,
original_score, start_inx+i, queue)))
for i in xrange(0, len(delta_inv_change_set)):
workers.append(Process(target=run_delta, args=(delta_inv_change_set[i],
delta_inv_type_set[i], change_set, type_set, search_config, original_config, bitcode,
original_score, start_inx+len(delta_change_set)+i, inv_queue)))
# run in parallel maximum cpu_no processes at a time
cpu_no = CPU_NO
# multiprocessing.cpu_count()
loop = int(len(workers)/cpu_no)
for i in xrange(0, loop):
for j in xrange(i*cpu_no, min((i+1)*cpu_no, len(workers))):
workers[j].start()
for j in xrange(i*cpu_no, min((i+1)*cpu_no, len(workers))):
workers[j].join()
while not queue.empty():
result = queue.get()
if min_score == -1 or result[1] < min_score:
pass_inx = result[0] - start_inx
inv_is_better = False
while not inv_queue.empty():
result = inv_queue.get()
if min_score == -1 or result[1] < min_score:
pass_inx = result[0] - start_inx - len(delta_change_set)
inv_is_better = True
search_counter += len(delta_change_set) + len(delta_inv_change_set)
#
# recursively search in pass delta or pass delta inverse
# right now keep searching for the first pass delta or
# pass delta inverse; later on we will integrate cost
# model here
#
if pass_inx != -1:
pass_change_set = delta_inv_change_set[pass_inx] if inv_is_better else delta_change_set[pass_inx]
pass_type_set = delta_inv_type_set[pass_inx] if inv_is_better else delta_type_set[pass_inx]
if len(pass_change_set) > 1:
# always reset to lowest precision
utilities.to_2nd_highest_precision(change_set, type_set)
dd_search_config(pass_change_set, pass_type_set, search_config, original_config, bitcode, original_score, 2)
else:
utilities.to_2nd_highest_precision(change_set, type_set)
utilities.to_highest_precision(pass_change_set, pass_type_set)
return
#
# stop searching when division greater than change set size
#
if div >= len(change_set):
utilities.to_highest_precision(change_set, type_set)
return
else:
dd_search_config(change_set, type_set, search_config, original_config, bitcode, original_score, 2*div)
return
