gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

New crossover operator

Browse Source
This commit is contained in:
Philippe Tillet
2014-09-14 04:29:29 -04:00
parent c4c8404d40
commit 4539164ed8
3 changed files with 57 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
autotune/python/autotune.py
View File

@@ -142,11 +142,8 @@ def do_tuning(config_fname, spec_fname, viennacl_root):
if __name__ == "__main__":
parser = argparse.ArgumentParser();
subparsers = parser.add_subparsers(dest='action')
print_devices_parser = subparsers.add_parser('list-devices', help='list the devices available')
tune_parser = subparsers.add_parser('tune', help='tune using a specific configuration file')
tune_parser.add_argument("--config", default="config.ini", required=False, type=str)
tune_parser.add_argument("--viennacl-root", default='', required=False, type=str)

59
autotune/python/genetic.py
View File

@@ -28,7 +28,20 @@ class GeneticOperators(object):
self.ParameterType = TemplateType.Parameters
self.build_template = build_template
self.cache = {}
self.indpb = 0.15
self.indpb = 0.1
def decode(self):
simd = 2**int(s[0:2], 2)
ls0 = 2**int(s[2:5],2)
ls1 = 2**int(s[5:8],2)
kL = 2**int(s[8:11],2)
mS = 2**int(s[11:14],2)
kS = 2**int(s[14:17],2)
nS = 2**int(s[17:20],2)
fetchA = s[20:22].count(1)
fetchB = s[22:24].count(1)
lf0 = 2**int(s[24:7],2)
return [simd, ls0, kL, ls1, mS, kS, nS, fetchA, fetchB, lf0, ls0*ls1/lf0]
def init(self):
while True:
@@ -41,33 +54,39 @@ class GeneticOperators(object):
if template.check(self.statement)==0 and occupancy_record.occupancy >= 10 :
return result
def crossover(self, ind1, ind2):
ind1[1], ind2[1] = ind2[1], ind1[1]
ind1[3], ind2[3] = ind2[3], ind1[3]
ind1[7], ind2[7] = ind2[7], ind1[7]
ind1[8], ind2[8] = ind2[8], ind1[8]
ind1[9], ind2[9] = ind2[9], ind1[9]
ind1[10], ind2[10] = ind2[10], ind1[10]
return ind1, ind2
def mutate(self, individual):
FetchingPolicy = vcl.atidlas.FetchingPolicy
while True:
new_individual = copy.deepcopy(individual)
for i in new_individual:
if random.random() < self.indpb:
coef = random.choice([1, 2])
coef = 2**np.random.geometric(0.8)
funs = [lambda x:max(1, x/coef), lambda x:x*coef]
F = random.choice(funs)
nF = funs[1] if F==funs[0] else funs[0]
#swapping-based mutations
def m0():
new_individual[1], new_individual[3] = new_individual[3], new_individual[1]
def m1():
new_individual[4], new_individual[6] = new_individual[6], new_individual[4]
def m2():
new_individual[9], new_individual[10] = new_individual[10], new_individual[9]
#value modification mutations
def m3():
new_individual[0] = random.choice(self.parameters[0])
new_individual[0] = F(new_individual[0])
def m4():
new_individual[1] = F(new_individual[1])
new_individual[9] = F(new_individual[9])
new_individual[1], new_individual[9] = F(new_individual[1]), F(new_individual[9])
def m5():
new_individual[2] = F(new_individual[2])
def m6():
new_individual[3] = F(new_individual[3])
new_individual[10] = F(new_individual[10])
new_individual[3], new_individual[10] = F(new_individual[3]), F(new_individual[10])
def m7():
new_individual[4] = F(new_individual[4])
def m8():
@@ -75,15 +94,25 @@ class GeneticOperators(object):
def m9():
new_individual[6] = F(new_individual[6])
def m10():
new_individual[7] = random.choice([x for x in self.parameters[7] if x!=new_individual[7]])
new_individual[7] = random.choice([x for x in self.parameters[7] if x!=individual[7]])
if new_individual[7]!=FetchingPolicy.FETCH_FROM_LOCAL and new_individual[8]!=FetchingPolicy.FETCH_FROM_LOCAL:
new_individual[9], new_individual[10] = 0, 0
else:
new_individual[9], new_individual[10] = new_individual[1], new_individual[3]
def m11():
new_individual[8] = random.choice([x for x in self.parameters[8] if x!=new_individual[8]])
new_individual[8] = random.choice([x for x in self.parameters[8] if x!=individual[8]])
if new_individual[7]!=FetchingPolicy.FETCH_FROM_LOCAL and new_individual[8]!=FetchingPolicy.FETCH_FROM_LOCAL:
new_individual[9], new_individual[10] = 0, 0
else:
new_individual[9], new_individual[10] = new_individual[1], new_individual[3]
def m12():
new_individual[9] = F(new_individual[9])
new_individual[10] = nF(new_individual[10])
new_individual[9], new_individual[10] = F(new_individual[9]), nF(new_individual[10])
def m13():
new_individual[10] = F(new_individual[10])
new_individual[9] = nF(new_individual[9])
new_individual[10], new_individual[9] = F(new_individual[10]), nF(new_individual[9])
def m14():
new_individual[1], new_individual[3] = F(new_individual[1]), nF(new_individual[3])
def m15():
new_individual[3], new_individual[1] = F(new_individual[3]), nF(new_individual[1])
random.choice([m0, m1, m2, m3, m4, m5, m6, m7, m8, m9, m10, m11, m12, m13])()
template = self.build_template(self.TemplateType.Parameters(*new_individual))
if not tools.skip(template, self.statement, self.device):

6
autotune/python/optimize.py
View File

@@ -46,11 +46,11 @@ def genetic(statement, context, TemplateType, build_template, parameter_names, a
toolbox.register("individual", deap.tools.initIterate, creator.Individual, GA.init)
toolbox.register("population", deap.tools.initRepeat, list, toolbox.individual)
toolbox.register("evaluate", GA.evaluate)
toolbox.register("mate", deap.tools.cxTwoPoint)
toolbox.register("mate", GA.crossover)
toolbox.register("mutate", GA.mutate)
toolbox.register("select", deap.tools.selBest)
pop = toolbox.population(n=50)
pop = toolbox.population(n=70)
hof = deap.tools.HallOfFame(1)
best_performer = lambda x: max([compute_perf(hof[0].fitness.values[0]) for t in x])
@@ -60,4 +60,4 @@ def genetic(statement, context, TemplateType, build_template, parameter_names, a
stats.register("max (" + perf_metric + ")", lambda x: max([compute_perf(hof[0].fitness.values[0]) for t in x]))
stats.register("profile ", lambda x: '(%s)'%','.join(map(str,hof[0])))
pop = eaMuPlusLambda(pop, toolbox, 50, 70, cxpb=0.2, mutpb=0.3, maxtime='5m0s', maxgen=500, halloffame=hof, compute_perf=compute_perf, perf_metric=perf_metric)
pop = eaMuPlusLambda(pop, toolbox, 70, 100, cxpb=0.2, mutpb=0.3, maxtime='5m0s', maxgen=1000, halloffame=hof, compute_perf=compute_perf, perf_metric=perf_metric)