class SweepTool(AbstractTool):
TOOL_NAME='sweep'
TOOL_SUMMARY='Generating timing measurements suitable for training a performance estimator.'
CACHE_FORMAT=[] # FIXME: Features are not cached.
def add_subparser(self, argparser):
super(SweepTool, self).add_subparser(argparser)
self.subparser.add_argument('primop', nargs='?', type=str, help='Which primitive operation to sweep.')
self.subparser.add_argument('-n', type=int, default=1, help='Number of data points to collect.')
self.subparser.add_argument('-o', '--output', type=str, default=None, help='Output file with Estimator training features (default: <primop>.features).')
self.subparser.add_argument('--seed', type=int, default=None, help='Set the PRNG seed for this sweep.')
self.subparser.add_argument('-l','--list', default=False, action='store_true', help="Don't actually run a sweep, just print which primops are supported,")
# FIXME: Add argument to limit scope of sweep
return self.subparser
def _run(self):
frame = TFFramework()
if self.args['list']:
print 'Supported Exemplars:'
for primop in frame.ExemplarRegistry:
print ' ',primop
return 0
elif self.args['primop'] is None:
self.subparser.error('A primop must be specified unless using --list')
primop_t = self.args['primop']
uas = UniformArgSampler()
primop_args = uas.sample(primop_t, n=self.args['n'], seed=self.args['seed'])
Exemplar = frame.ExemplarRegistry.lookup(primop_t)
self.features = Features()
for p_args in primop_args:
exemplar = Exemplar(p_args)
synthmodel = frame.SyntheticModel(exemplar)
frame.set_model(synthmodel)
profile = frame.get_timing(mode='inference', ops='native')
graph = frame.get_graph(mode='inference', scope='dynamic', ops='native')
id = graph.get_vertex_id_from_tf_name( exemplar.get_op_name() )
self.features.append( p_args, profile[id][0] )
if self.args['output'] is None:
self.args['output'] = str(primop_t)+'.features'
self.features.write(self.args['output'])
# FIXME: Features objects cannot be cached in the normal way.
#def _load(self, filename):
# feats = Features()
# feats.read(filename)
# return feats
#def _save(self, data, filename): # pylint: disable=W0221
# data.write(filename)
def _output(self):
print 'Output'
for args,time in zip(self.features.op_arguments, self.features.measurements):
print args,':',time
def test_io(self):
f = Features()
f.append([1,2,3], 10)
f.append([2,3,4], 11)
f.append([3,4,5], 12)
with in_temporary_directory():
f.write('test_file')
f2 = Features()
f2.read('test_file')
assert np.all(f.array==f2.array)