def _get_cluster(): named_device = device_properties_pb2.NamedDevice() named_device.name = '/GPU:0' named_device.properties.type = 'GPU' named_device.properties.environment['architecture'] = '4' cluster = gcluster.Cluster(devices=[named_device]) return cluster
def testVirtualCluster(self):
with ops.Graph().as_default() as g:
with ops.device('/device:GPU:0'):
a = random_ops.random_uniform(shape=[1024, 1024])
b = random_ops.random_uniform(shape=[1024, 1024])
c = a + b
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(c)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
device_properties = device_properties_pb2.DeviceProperties(
type='GPU',
frequency=1000,
num_cores=60,
environment={'architecture': '7'})
named_device = device_properties_pb2.NamedDevice(
properties=device_properties, name='/device:GPU:0')
grappler_cluster = cluster.Cluster(
disable_detailed_stats=False,
disable_timeline=False,
devices=[named_device])
op_perfs, run_time, _ = grappler_cluster.MeasureCosts(grappler_item)
self.assertEqual(run_time, 0.000545)
self.assertEqual(len(op_perfs), 15)
estimated_perf = grappler_cluster.EstimatePerformance(named_device)
self.assertEqual(7680.0, estimated_perf)
def testSupportDevices(self):
gpu_type = test_util.gpu_device_type()
gpu_name = test_util.gpu_device_name()
with ops.Graph().as_default() as g:
a = random_ops.random_uniform(shape=(2, 3))
b = random_ops.random_uniform(shape=(2, 3))
c = a + b
dims = math_ops.range(0, array_ops.rank(c), 1)
d = math_ops.reduce_sum(a, axis=dims)
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(d)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
device_properties = device_properties_pb2.DeviceProperties(
type=gpu_type, frequency=1000, num_cores=60)
named_gpu = device_properties_pb2.NamedDevice(
properties=device_properties, name=gpu_name)
device_properties = device_properties_pb2.DeviceProperties(
type='CPU', frequency=3000, num_cores=6)
named_cpu = device_properties_pb2.NamedDevice(
properties=device_properties, name='/CPU:0')
virtual_cluster = cluster.Cluster(devices=[named_cpu, named_gpu])
supported_dev = virtual_cluster.GetSupportedDevices(grappler_item)
self.assertEqual(supported_dev['add'], ['/CPU:0', gpu_name])
self.assertEqual(supported_dev['Sum'], ['/CPU:0', gpu_name])
self.assertEqual(supported_dev['range'], ['/CPU:0', gpu_name])
real_cluster = cluster.Cluster()
supported_dev = real_cluster.GetSupportedDevices(grappler_item)
if test.is_gpu_available():
self.assertEqual(supported_dev['add'], [
'/job:localhost/replica:0/task:0/device:CPU:0',
'/job:localhost/replica:0/task:0' + gpu_name
])
self.assertEqual(supported_dev['Sum'], [
'/job:localhost/replica:0/task:0/device:CPU:0',
'/job:localhost/replica:0/task:0' + gpu_name
])
# The axis tensor must reside on the host
self.assertEqual(
supported_dev['range'],
['/job:localhost/replica:0/task:0/device:CPU:0'])
else:
self.assertEqual(
supported_dev['add'],
['/job:localhost/replica:0/task:0/device:CPU:0'])
def get_cluster():
"""Grappler optimization configuration for GPU."""
named_device = device_properties_pb2.NamedDevice()
named_device.name = '/GPU:0'
named_device.properties.type = 'GPU'
named_device.properties.environment['architecture'] = '4'
cluster = gcluster.Cluster(devices=[named_device])
return cluster
def buildCluster(num_cpus=1, num_gpus=2):
devices = []
if num_gpus > 0:
device_properties = device_properties_pb2.DeviceProperties(
type='GPU',
vendor='NVidia',
model='GeForce GTX TITAN X',
frequency=1076,
num_cores=24,
environment={
'architecture': '5.2',
'cuda': '8000',
'cudnn': '6021'
},
num_registers=65536,
l1_cache_size=24576,
l2_cache_size=3145728,
shared_memory_size_per_multiprocessor=98304,
memory_size=12783648768,
bandwidth=336480000)
for i in range(num_gpus):
devices.append(
device_properties_pb2.NamedDevice(properties=device_properties,
name='/GPU:' + str(i)))
assert num_cpus > 0
device_properties = device_properties_pb2.DeviceProperties(
type='CPU',
frequency=1900,
num_cores=2,
l1_cache_size=32768,
l2_cache_size=262144,
l3_cache_size=3145728)
for i in range(num_cpus):
devices.append(
device_properties_pb2.NamedDevice(properties=device_properties,
name='/CPU:' + str(i)))
return cluster.Cluster(devices=devices)
def _buildCluster(num_cpus=1, num_gpus=1):
devices = []
if num_gpus > 0:
device_properties = device_properties_pb2.DeviceProperties(
type='GPU',
vendor='NVidia',
model='Tesla K40m',
frequency=745, #745 MHZ
num_cores=2888, # CUDA Cores
environment={
'architecture': '5.2',
'cuda': '10000',
'cudnn': '7031'
},
num_registers=65536,
l1_cache_size=65536, #64KB
l2_cache_size=1572864, #1.5 MB
shared_memory_size_per_multiprocessor=49152, #49152 bytes
memory_size=12884901888, # 12GB
bandwidth=288000000) #288 GBps
for i in range(num_gpus):
devices.append(
device_properties_pb2.NamedDevice(
properties=device_properties, name='/GPU:' + str(i)))
assert num_cpus > 0
device_properties = device_properties_pb2.DeviceProperties(
type='CPU',
frequency=2399,
num_cores=32,
l1_cache_size=32768,
l2_cache_size=262144,
l3_cache_size=20971520)
for i in range(num_cpus):
devices.append(
device_properties_pb2.NamedDevice(properties=device_properties,
name='/CPU:' + str(i)))
return cluster.Cluster(devices=devices)
def build_cluster():
devices = []
device_properties = device_properties_pb2.DeviceProperties(
type='CPU',
frequency=2000,
num_cores=12,
l1_cache_size=32768,
l2_cache_size=262144,
l3_cache_size=30720*1024)
for i in range(2):
devices.append(
device_properties_pb2.NamedDevice(
properties=device_properties, name='/CPU:' + str(i)))
return cluster.Cluster(devices=devices)
def testVirtualCluster(self):
with ops.Graph().as_default() as g:
a = random_ops.random_uniform(shape=())
b = random_ops.random_uniform(shape=())
c = a + b
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(c)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
device_properties = device_properties_pb2.DeviceProperties(
type='GPU', environment={
'architecture': '7'
})
named_device = device_properties_pb2.NamedDevice(
properties=device_properties, name='/GPU:0')
grappler_cluster = cluster.Cluster(devices=[named_device])
op_perfs, run_time, _ = grappler_cluster.MeasureCosts(grappler_item)
self.assertGreater(run_time, 0)
self.assertEqual(len(op_perfs), 15)
def testSupportDevices(self):
with ops.Graph().as_default() as g:
a = random_ops.random_uniform(shape=(2, 3))
b = random_ops.random_uniform(shape=(2, 3))
c = a + b
dims = math_ops.range(0, array_ops.rank(c), 1)
d = math_ops.reduce_sum(a, axis=dims)
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(d)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
device_properties = device_properties_pb2.DeviceProperties(
type='GPU', frequency=1000, num_cores=60)
named_gpu = device_properties_pb2.NamedDevice(
properties=device_properties, name='/GPU:0')
device_properties = device_properties_pb2.DeviceProperties(
type='CPU', frequency=3000, num_cores=6)
named_cpu = device_properties_pb2.NamedDevice(
properties=device_properties, name='/CPU:0')
virtual_cluster = cluster.Cluster(devices=[named_cpu, named_gpu])
supported_dev = virtual_cluster.GetSupportedDevices(grappler_item)
self.assertEqual(supported_dev['add'], ['/CPU:0', '/GPU:0'])
self.assertEqual(supported_dev['Sum'], ['/CPU:0', '/GPU:0'])
self.assertEqual(supported_dev['range'], ['/CPU:0', '/GPU:0'])
real_cluster = cluster.Cluster()
supported_dev = real_cluster.GetSupportedDevices(grappler_item)
#NCL 18.04 -- Hack to account for possible XLA devices
if test.is_gpu_available():
add_devices = [
d for d in supported_dev['add']
if not d.split(':')[-2].startswith('XLA')
]
self.assertEqual(add_devices, [
'/job:localhost/replica:0/task:0/device:CPU:0',
'/job:localhost/replica:0/task:0/device:GPU:0'
])
Sum_devices = [
d for d in supported_dev['Sum']
if not d.split(':')[-2].startswith('XLA')
]
self.assertEqual(Sum_devices, [
'/job:localhost/replica:0/task:0/device:CPU:0',
'/job:localhost/replica:0/task:0/device:GPU:0'
])
# The axis tensor must reside on the host
range_devices = [
d for d in supported_dev['range']
if not d.split(':')[-2].startswith('XLA')
]
self.assertEqual(
range_devices,
['/job:localhost/replica:0/task:0/device:CPU:0'])
else:
add_devices = [
d for d in supported_dev['add']
if not d.split(':')[-2].startswith('XLA')
]
self.assertEqual(
add_devices,
['/job:localhost/replica:0/task:0/device:CPU:0'])
