Tuning: Android UI improvement
This commit is contained in:
Philippe Tillet
@@ -12,14 +12,13 @@ from external.deap import tools as deap_tools
|
||||
from numpy import cumsum
|
||||
|
||||
import tools
|
||||
from tools import profile_execution_failure
|
||||
|
||||
fetch_types = [sc.templates.fetching_policy_type.FETCH_FROM_LOCAL,
|
||||
sc.templates.fetching_policy_type.FETCH_FROM_LOCAL,
|
||||
sc.templates.fetching_policy_type.FETCH_FROM_LOCAL,
|
||||
sc.templates.fetching_policy_type.FETCH_FROM_LOCAL]
|
||||
|
||||
to_catch = (sc.OperationNotSupported, sc.OclLaunchOutOfResources, sc.CudaLaunchOutOfResources, sc.MemObjectAllocationFailure, sc.InvalidWorkGroupSize, sc.OutOfHostMemory, sc.InvalidValue)
|
||||
|
||||
def exhaustive(template, sizes, context):
|
||||
tree, _ = tools.tree_of(template, sizes, context)
|
||||
metric = tools.metric_of(template)
|
||||
@@ -35,127 +34,137 @@ def exhaustive(template, sizes, context):
|
||||
time = tools.benchmark(template, parameters, tree)
|
||||
if not best or time < best[1]:
|
||||
best = parameters, time
|
||||
except to_catch:
|
||||
except profile_execution_failure:
|
||||
pass
|
||||
if best:
|
||||
stdout.write('%.2f %% | Best %.2f [ for %s ]\r'%(float(idx*100)/len(ranges),metric(sizes, best[1]), best[0]))
|
||||
return best[0]
|
||||
|
||||
|
||||
def genetic(template, sizes, context, naccept=200, niter = 1000, cxpb=0.4, mutpb=0.4, popsize = 10, initializer = None, prior = None):
|
||||
tree, _ = tools.tree_of(template, sizes, context)
|
||||
metric = tools.metric_of(template)
|
||||
genetic_infos = tools.genetic_infos_of(template)
|
||||
nbits = genetic_infos['nbits']
|
||||
offsets = cumsum([0] + nbits)
|
||||
|
||||
def bin2gray(A):
|
||||
g = [int(A[0])]
|
||||
for i in range(1, len(A)):
|
||||
g += [int(A[i-1] != A[i])]
|
||||
return g
|
||||
class GeneticOptimizer:
|
||||
|
||||
def gray2int(A):
|
||||
b = [A[0]]
|
||||
for i in range(1, len(A)):
|
||||
b += [int(b[i-1] != A[i])]
|
||||
return int(''.join(map(str,b)), 2)
|
||||
|
||||
def encode(genome):
|
||||
encoded = [bin2gray(bin(x)[2:].zfill(nb)) for x, nb in zip(genome, nbits)]
|
||||
return sum(encoded, [])
|
||||
def __init__(self, logger, naccept=500, niter=1000, cxpb=.4, mutpb=.4, popsize=10):
|
||||
self.logger = logger
|
||||
self.naccept = naccept
|
||||
self.niter = niter
|
||||
self.cxpb = cxpb
|
||||
self.mutpb = mutpb
|
||||
self.popsize = popsize
|
||||
|
||||
def decode(genome):
|
||||
result = []
|
||||
for off1,off2 in zip(offsets[:-1],offsets[1:]):
|
||||
result += [gray2int(genome[off1:off2])]
|
||||
result = [fetch_types[x] if i in genetic_infos['categorical'] else 2**x for i,x in enumerate(result)]
|
||||
return result
|
||||
def run(self, template, sizes, context, initializer = None, prior = None):
|
||||
tree, _ = tools.tree_of(template, sizes, context)
|
||||
metric = tools.metric_of(template)
|
||||
genetic_infos = tools.genetic_infos_of(template)
|
||||
nbits = genetic_infos['nbits']
|
||||
offsets = cumsum([0] + nbits)
|
||||
|
||||
def evaluate(genome):
|
||||
idx = tuple(genome)
|
||||
if idx not in cache:
|
||||
cache[idx] = tools.benchmark(template, decode(genome), tree)
|
||||
return cache[idx],
|
||||
def bin2gray(A):
|
||||
g = [int(A[0])]
|
||||
for i in range(1, len(A)):
|
||||
g += [int(A[i-1] != A[i])]
|
||||
return g
|
||||
|
||||
cache = {}
|
||||
hof = deap_tools.HallOfFame(1)
|
||||
|
||||
creator.create("FitnessMin", base.Fitness, weights=(-1.0,))
|
||||
creator.create("Individual", list, fitness=creator.FitnessMin)
|
||||
|
||||
toolbox = base.Toolbox()
|
||||
toolbox.register("evaluate", evaluate)
|
||||
toolbox.register("mate", deap_tools.cxTwoPoint)
|
||||
toolbox.register("mutate", deap_tools.mutFlipBit)
|
||||
toolbox.register("select", deap_tools.selNSGA2)
|
||||
|
||||
#Initialization
|
||||
if initializer is None:
|
||||
initializer = ([random.randint(0, 2**x) for x in nbits] for i in iter(int,1))
|
||||
population = []
|
||||
|
||||
genome = encode(prior if prior else list(initializer.next()))
|
||||
while len(population) < popsize:
|
||||
individual = creator.Individual(genome)
|
||||
try:
|
||||
individual.fitness.values = toolbox.evaluate(genome)
|
||||
population += [individual]
|
||||
except to_catch:
|
||||
pass
|
||||
genome = encode(list(initializer.next()))
|
||||
hof.update(population)
|
||||
|
||||
x = []
|
||||
y = []
|
||||
it = 0
|
||||
|
||||
while len(cache) < naccept and it<niter:
|
||||
pad = len(cache) - len(x)
|
||||
x += [len(cache)]*pad
|
||||
y += [metric(sizes, hof[0].fitness.values[0])]*pad
|
||||
def gray2int(A):
|
||||
b = [A[0]]
|
||||
for i in range(1, len(A)):
|
||||
b += [int(b[i-1] != A[i])]
|
||||
return int(''.join(map(str,b)), 2)
|
||||
|
||||
offspring = []
|
||||
while len(offspring) < popsize:
|
||||
try:
|
||||
op_choice = random.random()
|
||||
#Cross-over
|
||||
if op_choice < cxpb:
|
||||
ind1, ind2 = map(toolbox.clone, random.sample(population, 2))
|
||||
ind1, ind2 = toolbox.mate(ind1, ind2)
|
||||
ind = ind1
|
||||
toolbox.evaluate(ind)
|
||||
offspring += [ind]
|
||||
#Mutation
|
||||
elif op_choice < cxpb + mutpb:
|
||||
ind = toolbox.clone(random.choice(population))
|
||||
ind, = toolbox.mutate(ind, 1.0/offsets[-1])
|
||||
toolbox.evaluate(ind)
|
||||
offspring += [ind]
|
||||
#Reproduction
|
||||
else:
|
||||
offspring += [random.choice(population)]
|
||||
except to_catch:
|
||||
pass
|
||||
|
||||
|
||||
#Update fitnesses
|
||||
fitnesses = toolbox.map(toolbox.evaluate, offspring)
|
||||
for ind, fit in zip(offspring, fitnesses):
|
||||
ind.fitness.values = fit
|
||||
def encode(genome):
|
||||
encoded = [bin2gray(bin(x)[2:].zfill(nb)) for x, nb in zip(genome, nbits)]
|
||||
return sum(encoded, [])
|
||||
|
||||
#Update population
|
||||
population[:] = toolbox.select(population + offspring, popsize)
|
||||
def decode(genome):
|
||||
result = []
|
||||
for off1,off2 in zip(offsets[:-1],offsets[1:]):
|
||||
result += [gray2int(genome[off1:off2])]
|
||||
result = [fetch_types[x] if i in genetic_infos['categorical'] else 2**x for i,x in enumerate(result)]
|
||||
return result
|
||||
|
||||
def evaluate(genome):
|
||||
idx = tuple(genome)
|
||||
if idx not in cache:
|
||||
cache[idx] = tools.benchmark(template, decode(genome), tree)
|
||||
return cache[idx],
|
||||
|
||||
cache = {}
|
||||
hof = deap_tools.HallOfFame(1)
|
||||
|
||||
creator.create("FitnessMin", base.Fitness, weights=(-1.0,))
|
||||
creator.create("Individual", list, fitness=creator.FitnessMin)
|
||||
|
||||
toolbox = base.Toolbox()
|
||||
toolbox.register("evaluate", evaluate)
|
||||
toolbox.register("mate", deap_tools.cxTwoPoint)
|
||||
toolbox.register("mutate", deap_tools.mutFlipBit)
|
||||
toolbox.register("select", deap_tools.selNSGA2)
|
||||
|
||||
#Initialization
|
||||
if initializer is None:
|
||||
initializer = ([random.randint(0, 2**x) for x in nbits] for i in iter(int,1))
|
||||
population = []
|
||||
|
||||
genome = encode(prior if prior else list(initializer.next()))
|
||||
while len(population) < self.popsize:
|
||||
individual = creator.Individual(genome)
|
||||
try:
|
||||
individual.fitness.values = toolbox.evaluate(genome)
|
||||
population += [individual]
|
||||
except profile_execution_failure:
|
||||
pass
|
||||
genome = encode(list(initializer.next()))
|
||||
hof.update(population)
|
||||
|
||||
optimal = '(%s)'%','.join(map(str,decode(hof[0])))
|
||||
stdout.write('Iter %d | %d evaluated | Best %.2f [ for %s ]\r'%(it, x[-1], y[-1], optimal))
|
||||
stdout.flush()
|
||||
it += 1
|
||||
stdout.write('\n')
|
||||
x = []
|
||||
y = []
|
||||
it = 0
|
||||
|
||||
while len(cache) < self.naccept and it<self.niter:
|
||||
pad = len(cache) - len(x)
|
||||
x += [len(cache)]*pad
|
||||
y += [metric(sizes, hof[0].fitness.values[0])]*pad
|
||||
|
||||
offspring = []
|
||||
while len(offspring) < self.popsize:
|
||||
try:
|
||||
op_choice = random.random()
|
||||
#Cross-over
|
||||
if op_choice < self.cxpb:
|
||||
ind1, ind2 = map(toolbox.clone, random.sample(population, 2))
|
||||
ind1, ind2 = toolbox.mate(ind1, ind2)
|
||||
ind = ind1
|
||||
toolbox.evaluate(ind)
|
||||
offspring += [ind]
|
||||
#Mutation
|
||||
elif op_choice < self.cxpb + self.mutpb:
|
||||
ind = toolbox.clone(random.choice(population))
|
||||
ind, = toolbox.mutate(ind, 1.0/offsets[-1])
|
||||
toolbox.evaluate(ind)
|
||||
offspring += [ind]
|
||||
#Reproduction
|
||||
else:
|
||||
offspring += [random.choice(population)]
|
||||
except profile_execution_failure:
|
||||
pass
|
||||
|
||||
return tuple(decode(hof[0])), x, y
|
||||
|
||||
|
||||
#Update fitnesses
|
||||
fitnesses = toolbox.map(toolbox.evaluate, offspring)
|
||||
for ind, fit in zip(offspring, fitnesses):
|
||||
ind.fitness.values = fit
|
||||
|
||||
#Update population
|
||||
population[:] = toolbox.select(population + offspring, self.popsize)
|
||||
hof.update(population)
|
||||
|
||||
optimal = '(%s)'%','.join(map(str,decode(hof[0])))
|
||||
stdout.write('Iter %d | %d evaluated | Best %.2f [ for %s ]\r'%(it, x[-1], y[-1], optimal))
|
||||
stdout.flush()
|
||||
it += 1
|
||||
stdout.write('\n')
|
||||
|
||||
return tuple(decode(hof[0])), x, y
|
||||
|
||||
def is_local_optimum(parameters, template, sizes, context):
|
||||
tree, _ = tools.tree_of(template, sizes, context)
|
||||
genetic_infos = tools.genetic_infos_of(template)
|
||||
@@ -174,7 +183,7 @@ def is_local_optimum(parameters, template, sizes, context):
|
||||
#Evaluate the provided parameters guess
|
||||
try:
|
||||
reference = tools.benchmark(template, parameters, tree)
|
||||
except to_catch:
|
||||
except profile_execution_failure:
|
||||
return False
|
||||
|
||||
#Latency bound -- ignore
|
||||
@@ -191,7 +200,7 @@ def is_local_optimum(parameters, template, sizes, context):
|
||||
time = tools.benchmark(template, x, tree)
|
||||
if time/reference < .97:
|
||||
return False
|
||||
except to_catch:
|
||||
except profile_execution_failure:
|
||||
pass
|
||||
return True
|
||||
|
||||
|
||||
Reference in New Issue
Block a user