def eval_classifier_instance(job, db, classifier, instance,
err_fn, training):
# Get relevant values from instance.
oracle = instance.get_string_value(instance.class_index)
scenario = instance.get_string_value(0)
# Get default value.
try:
baseline = training.default
except AttributeError:
training.default = get_one_r(db, training)
baseline = training.default
# Classify instance, and convert to params ID.
prof.start()
value = classifier.classify_instance(instance)
elapsed = prof.elapsed()
attr = instance.dataset.attribute(instance.class_index)
predicted = attr.value(value)
correct = 1 if predicted == oracle else 0
if correct:
illegal = 0
refused = 0
performance, speedup = perf_fn(db, scenario, predicted,
oracle, baseline)
else:
# Determine if predicted workgroup size is valid or not. A
# valid predicted is one which is within the max_wgsize for
# that particular instance.
max_wgsize_attr = instance.dataset.attribute_by_name("kern_max_wg_size")
max_wgsize_attr_index = max_wgsize_attr.index
max_wgsize = int(instance.get_value(max_wgsize_attr_index))
wg_c, wg_r = unhash_params(predicted)
illegal = 0 if wg_c * wg_r < max_wgsize else 1
if illegal:
refused = 0
else:
try:
db.runtime(scenario, predicted)
refused = 0
except lab.db.Error:
refused = 1
if illegal or refused:
predicted = err_fn(instance, max_wgsize, wg_c, wg_r, baseline)
performance, speedup = perf_fn(db, scenario, predicted,
oracle, baseline)
db.add_classification_result(job, classifier,
err_fn, training, scenario,
oracle, predicted, baseline, correct, illegal,
refused, performance, speedup, elapsed)
def reshape(db, scenario, max_wgsize, wg_c, wg_r):
W_legal = db.W_legal(scenario)
min_distance = float("inf")
min_param = None
# Find the *legal* parameter which is closest to the predicted by
# calculating the distance to each and returning the smallest.
for param in W_legal:
p_c, p_r = unhash_params(param)
distance = math.sqrt((wg_c - p_c) ** 2 + (wg_r - p_r) ** 2)
if distance < min_distance:
min_distance = distance
min_param = param
return min_param
def run_job(i, n, wgsize, program, args):
wg_c, wg_r = unhash_params(wgsize)
# Set environment variable.
os.environ["OMNITUNE_OFFLINE_TRAINING"] = "1"
os.environ["OMNITUNE_STENCIL_WG_C"] = str(wg_c)
os.environ["OMNITUNE_STENCIL_WG_R"] = str(wg_r)
fs.cd(fs.path(experiment.EXAMPLES_BUILD, program))
cmd_str = "./{} {}".format(program, args.rstrip())
cmd = cmd_str.split()
io.info(i, "of", n, " - ", wgsize, "COMMAND:", io.colourise(io.Colours.RED, cmd_str))
ret, _, _ = system.run(cmd, stdout=system.STDOUT, stderr=system.STDERR)
if ret:
print(ret, wgsize, program, args, sep="\t", file=errlog)
else:
print(ret, wgsize, program, args, sep="\t", file=runlog)
def eval_speedup_regressors(db, classifiers, baseline, rank_fn,
table, job, training, testing):
maxwgsize_index = testing.attribute_by_name("kern_max_wg_size").index
wg_c_index = testing.attribute_by_name("wg_c").index
wg_r_index = testing.attribute_by_name("wg_r").index
insert_str = ("INSERT INTO {} VALUES "
"(?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)".format(table))
for classifier in classifiers:
meta = Classifier(classifier)
prof.start("train classifier")
meta.build_classifier(training)
prof.stop("train classifier")
basename = ml.classifier_basename(classifier.classname)
classifier_id = db.classifier_id(classifier)
io.debug(job, basename, testing.num_instances)
scenarios = set([instance.get_string_value(0)
for instance in testing])
instances = zip(scenarios, [
(instance for instance in testing if
instance.get_string_value(0) == scenario).next()
for scenario in scenarios
])
for scenario,instance in instances:
maxwgsize = int(instance.get_value(maxwgsize_index))
wlegal = space.enumerate_wlegal_params(maxwgsize)
predictions = []
elapsed = 0
for params in wlegal:
wg_c, wg_r = unhash_params(params)
instance.set_value(wg_c_index, wg_c)
instance.set_value(wg_r_index, wg_r)
# Predict the speedup for a particular set of
# parameters.
prof.start()
predicted = meta.classify_instance(instance)
elapsed += prof.elapsed()
predictions.append((params, predicted))
# Rank the predictions from highest to lowest speedup.
predictions = sorted(predictions, key=lambda x: x[1], reverse=True)
row = db.execute(
"SELECT "
" oracle_param,"
" ("
" SELECT mean FROM runtime_stats "
" WHERE scenario=? AND params=?"
" ) * 1.0 / oracle_runtime AS oracle_speedup,"
" worst_runtime / oracle_runtime AS actual_range "
"FROM scenario_stats "
"WHERE scenario=?",
(scenario,baseline,scenario)).fetchone()
actual = row[:2]
predicted_range = predictions[-1][1] - predictions[0][1]
actual_range = row[2] - row[1]
num_attempts = 1
while True:
predicted = predictions.pop(0)
try:
speedup = db.speedup(scenario, baseline, predicted[0])
perf = db.perf(scenario, predicted[0])
try:
speedup_he = self.speedup(scenario, HE_PARAM, predicted[0])
except:
speedup_he = None
try:
speedup_mo = self.speedup(scenario, MO_PARAM, predicted[0])
except:
speedup_mo = None
db.execute(insert_str,
(job, classifier_id, scenario, actual[0],
actual[1], predicted[0], predicted[1],
actual_range, predicted_range,
num_attempts,
1 if predicted[0] == actual[0] else 0,
perf, speedup, speedup_he, speedup_mo, elapsed))
break
except _db.MissingDataError:
num_attempts += 1
pass
db.commit()
