def setup(self):
from neon.backends.mgpu import MGPU, MGPUTensor
# this code gets called prior to each test
try:
self.be = MGPU(rng_seed=0, num_dev=2)
except AssertionError:
# likely that only one GPU device is available
self.be = MGPU(rng_seed=0, num_dev=1)
self.gpt = MGPUTensor
def setup(self):
from neon.backends.mgpu import MGPU, MGPUTensor
# this code gets called prior to each test
try:
self.be = MGPU(rng_seed=0, num_dev=2)
except AssertionError:
# likely that only one GPU device is available
self.be = MGPU(rng_seed=0, num_dev=1)
self.gpt = MGPUTensor
class TestGPU(object):
def setup(self):
from neon.backends.mgpu import MGPU, MGPUTensor
# this code gets called prior to each test
try:
self.be = MGPU(rng_seed=0, num_dev=2)
except AssertionError:
# likely that only one GPU device is available
self.be = MGPU(rng_seed=0, num_dev=1)
self.gpt = MGPUTensor
def reduction_test(self):
nr = self.be.num_dev
if nr == 1: # This shouldn't be supported
return
# create a numpy array as the test-bed
asize = 9
# round up to the nearest multiple of num_dev
bsize = -(-asize // nr) * nr
h_a = np.random.randn(asize * nr).reshape(
(nr, asize)).astype(self.be.default_dtype)
h_result = np.sum(h_a, axis=0, keepdims=True)
d_a = self.be.empty((1, asize))
u_a = self.be.empty((1, bsize))
self.be.scatter(h_a, d_a)
self.be.reduce(d_a, u_a)
print(h_result)
print(d_a.tlist[0].asnumpyarray())
for i in range(nr):
np.testing.assert_allclose(d_a.tlist[i].asnumpyarray(),
h_result,
atol=1e-6,
rtol=0)
def memset_test(self):
# create a numpy array as the test-bed
asize = 9
h_result = np.zeros((1, asize))
d_a = self.be.zeros((1, asize))
for i in range(self.be.num_dev):
np.testing.assert_allclose(d_a.tlist[i].asnumpyarray(),
h_result,
atol=1e-6,
rtol=0)
def frag2rep_test(self):
nr = self.be.num_dev
if nr == 1: # This shouldn't be supported
return
np.random.seed(0)
# create a numpy array as the test-bed
(rows, cols) = (24, 128)
indim = rows * cols
odim = indim * nr
# h_frags has the data in the order we expect on the device
h_frags_t = np.random.randn(odim).reshape(
(nr * cols, rows)).astype(self.be.default_dtype)
h_frags = h_frags_t.transpose().astype(self.be.default_dtype,
order='C')
d_frags = self.be.empty((rows, cols))
d_frags_t = self.be.empty((cols, rows))
d_reps = self.be.empty((rows, cols * nr))
d_reps_t = self.be.empty((cols * nr, rows))
self.be.scatter(h_frags_t, d_frags_t)
self.be.transpose(d_frags_t, d_frags)
np.testing.assert_allclose(d_frags.asnumpyarray(),
h_frags,
atol=1e-5,
rtol=0)
self.be.fragment_to_replica(d_frags_t, d_reps_t)
self.be.transpose(d_reps_t, d_reps)
for i in range(nr):
np.testing.assert_allclose(d_frags.asnumpyarray(),
d_reps.tlist[i].asnumpyarray(),
atol=1e-5,
rtol=0)
print("Frag2Rep OK")
d_frags_t.fill(0)
self.be.replica_to_fragment(d_reps_t, d_frags_t)
self.be.transpose(d_frags_t, d_frags)
for i in range(nr):
np.testing.assert_allclose(d_frags.asnumpyarray(),
d_reps.tlist[i].asnumpyarray(),
atol=1e-5,
rtol=0)
print("Rep2Frag OK")
class TestGPU(object):
def setup(self):
from neon.backends.mgpu import MGPU, MGPUTensor
# this code gets called prior to each test
try:
self.be = MGPU(rng_seed=0, num_dev=2)
except AssertionError:
# likely that only one GPU device is available
self.be = MGPU(rng_seed=0, num_dev=1)
self.gpt = MGPUTensor
def reduction_test(self):
nr = self.be.num_dev
if nr == 1: # This shouldn't be supported
return
# create a numpy array as the test-bed
asize = 9
# round up to the nearest multiple of num_dev
bsize = -(-asize // nr) * nr
h_a = np.random.randn(asize * nr).reshape(
(nr, asize)).astype(self.be.default_dtype)
h_result = np.sum(h_a, axis=0, keepdims=True)
d_a = self.be.empty((1, asize))
u_a = self.be.empty((1, bsize))
self.be.scatter(h_a, d_a)
self.be.reduce(d_a, u_a)
print(h_result)
print(d_a.tlist[0].asnumpyarray())
for i in range(nr):
np.testing.assert_allclose(d_a.tlist[i].asnumpyarray(),
h_result, atol=1e-6, rtol=0)
def memset_test(self):
# create a numpy array as the test-bed
asize = 9
h_result = np.zeros((1, asize))
d_a = self.be.zeros((1, asize))
for i in range(self.be.num_dev):
np.testing.assert_allclose(d_a.tlist[i].asnumpyarray(),
h_result, atol=1e-6, rtol=0)
def frag2rep_test(self):
nr = self.be.num_dev
if nr == 1: # This shouldn't be supported
return
np.random.seed(0)
# create a numpy array as the test-bed
(rows, cols) = (24, 128)
indim = rows * cols
odim = indim * nr
# h_frags has the data in the order we expect on the device
h_frags_t = np.random.randn(odim).reshape(
(nr * cols, rows)).astype(self.be.default_dtype)
h_frags = h_frags_t.transpose().astype(
self.be.default_dtype, order='C')
d_frags = self.be.empty((rows, cols))
d_frags_t = self.be.empty((cols, rows))
d_reps = self.be.empty((rows, cols * nr))
d_reps_t = self.be.empty((cols * nr, rows))
self.be.scatter(h_frags_t, d_frags_t)
self.be.transpose(d_frags_t, d_frags)
np.testing.assert_allclose(d_frags.asnumpyarray(),
h_frags, atol=1e-5, rtol=0)
self.be.fragment_to_replica(d_frags_t, d_reps_t)
self.be.transpose(d_reps_t, d_reps)
for i in range(nr):
np.testing.assert_allclose(d_frags.asnumpyarray(),
d_reps.tlist[i].asnumpyarray(),
atol=1e-5, rtol=0)
print("Frag2Rep OK")
d_frags_t.fill(0)
self.be.replica_to_fragment(d_reps_t, d_frags_t)
self.be.transpose(d_frags_t, d_frags)
for i in range(nr):
np.testing.assert_allclose(d_frags.asnumpyarray(),
d_reps.tlist[i].asnumpyarray(),
atol=1e-5, rtol=0)
print("Rep2Frag OK")
def gen_backend(model=None,
gpu=None,
nrv=False,
flexpoint=False,
rng_seed=None,
numerr_handling=None,
half=False,
stochastic_round=0,
device_id=None):
"""
Construct and return a backend instance of the appropriate type based on
the arguments given. With no parameters, a single CPU core, float32
backend is returned.
Arguments:
model (neon.models.model.Model): The instantiated model upon which we
will utilize this backend.
gpu (string, optional): Attempt to utilize a CUDA capable GPU if
installed in the system. Defaults to None which
implies a CPU based backend. If 'cudanet',
utilize a cuda-convnet2 based backed, which
supports Kepler and Maxwell GPUs with single
precision. If 'nervanagpu', attempt to utilize
the NervanaGPU Maxwell backend with float16 and
float32 support.
nrv (bool, optional): If True, attempt to utilize the Nervana Engine
for computation (must be installed on the
system). Defaults to False which implies a CPU
based backend.
rng_seed (numeric, optional): Set this to a numeric value which can be
used to seed the random number generator
of the instantiated backend. Defaults to
None, which doesn't explicitly seed (so
each run will be different)
stochastic_round (numeric, optional): Only affects the max backend. If
1, perform stochastic rounding.
If 0, round to nearest.
numerr_handling (dict, optional): Dictate how numeric errors are
displayed and handled. The keys and
values permissible for this dict
match that seen in numpy.seterr.
If set to None (the default),
behavior is equivalent to
{'all': 'warn'}
device_id (numeric, optional): Set this to a numeric value which can be
used to select which device to run the
process on
Returns:
Backend: newly constructed backend instance of the specifed type.
Notes:
* Attempts to construct a GPU instance without a CUDA capable card or
without cudanet or nervanagpu package installed will cause the
program to display an error message and exit.
* Attempts to construct a parallel instance without mpi4py installed
will cause the program to display an error message and exit.
* The returned backend will still need to call its par.init_model()
at some point after the model has been linked, in order for parallel
training to proceed.
"""
logger = logging.getLogger(__name__)
gpuflag = False
if gpu is not None:
gpu = gpu.lower()
if sys.platform.startswith("linux"):
gpuflag = (os.system("nvcc --version > /dev/null 2>&1") == 0)
elif sys.platform.startswith("darwin"):
gpuflag = (
os.system("kextstat | grep -i cuda > /dev/null 2>&1") == 0)
if gpuflag and gpu == 'cudanet':
try:
import cudanet # noqa
from neon.backends.cc2 import GPU
be_name = 'Cudanet'
be = GPU(rng_seed=rng_seed, device_id=device_id)
except ImportError:
raise RuntimeError("cudanet not found, can't run via GPU")
elif gpuflag and gpu.startswith('nervanagpu'):
try:
import nervanagpu # noqa
try:
be_name = 'NervanaGPU'
if gpu == 'nervanagpu':
device_id = 0 if device_id is None else device_id[0]
from neon.backends.gpu import GPU
be = GPU(rng_seed=rng_seed,
stochastic_round=stochastic_round,
device_id=device_id)
else:
from neon.backends.mgpu import MGPU
try:
num_dev = int(gpu.strip('nervanagpu'))
except (ValueError):
raise ValueError("invalid number of GPUs" +
" specified")
if not device_id:
device_id = range(num_dev)
if len(device_id) != num_dev:
raise RuntimeError(
"Incorrect number of devices"
" specified ", device_id, num_dev)
be = MGPU(rng_seed=rng_seed,
stochastic_round=stochastic_round,
device_id=device_id,
num_dev=num_dev)
except ImportError:
logger.warning("pycuda error, can't run via GPU")
gpuflag = False
except ImportError:
logger.warning("nervanagpu not found, can't run via GPU")
gpuflag = False
if gpuflag is False:
raise RuntimeError("Can't find CUDA capable GPU")
elif nrv:
nrv = False
try:
from umd.nrv_backend import NRVBackend
nrv = True
except ImportError:
logger.warning("Nervana Engine system software not found")
if flexpoint:
logger.warning("Flexpoint(TM) backend not currently available")
if nrv:
be_name = 'NRV'
be = NRVBackend(rng_seed=rng_seed,
seterr_handling=numerr_handling,
device_id=device_id)
elif not gpuflag:
be_name = 'CPU'
be = CPU(rng_seed=rng_seed, seterr_handling=numerr_handling)
logger.info("{} backend, RNG seed: {}, numerr: {}".format(
be_name, rng_seed, numerr_handling))
return be
