def run_tests():
bb = 1
cc = 1
ee = 1000
k_num = 10
a1 = np.random.random((bb, ee, cc))
a2 = np.random.random((bb, ee, cc))
for k_ee in range(13, 14):
b = np.random.random((k_num, k_ee, cc))
for md in ['valid', 'same', 'full']:
for orientation in ['as-is', 'flipped']:
import time
t1 = time.time()
y1 = reference(a1, b, md, orientation, 'NEC')
t2 = time.time()
y2 = reference(a2, b, md, orientation, 'NEC')
op = conv_1d(a1, b, mode=md, kernel_orientation=orientation, data_format='NEC')
# result1 =
if ovl.cuda_enabled:
assert np.allclose(ovl.evaluate(op, target_language='cuda'), y1)
# for d in range(1):
a1[:] = a2[:]
if ovl.cuda_enabled:
assert np.allclose(ovl.evaluate(op, target_language='cuda'), y2)
res, prof = ovl.profile(op, target_language='cuda', profiling_iterations=100, opt_level=3)
ovl.logger.debug(k_ee, md, orientation, (t2 - t1) * 1000, np.min(list(prof.values())[0]))
def countTrianglesCPU(startEdge, fromVertex, toVertex):
"""Count the triangles on the CPU.
The array toVertex is a flattened list of lists structure, where startEdge encodes the start indices of the lists.
:param startEdge: Indices into toVertex where edges start.
:type startEdge: list.
:param fromVertex: The from-vertex of each edge.
:type fromVertex: list.
:param toVertex: The to-vertex of each edge.
:type toVertex: list.
:return: Triangle count of graph.
:Examples:
.. doctest::
>>> from opveclib.examples.test_graph import loadGraphFromTextFile, writeExampleGraphToTextFile, countTrianglesCPU
>>> tmpName = "/tmp/v7e20.txt"
>>> writeExampleGraphToTextFile(tmpName)
>>> startEdge, fromVertex, toVertex = loadGraphFromTextFile(tmpName)
>>> countTrianglesCPU(startEdge, fromVertex, toVertex)
3
"""
count = ovl.evaluate(graph_triangle_count(startEdge, fromVertex, toVertex),
target_language='cpp')
return np.sum(count, axis=0, dtype=np.uint64)
def triangles(startEdge, fromVertex, toVertex, target_language='cpp'):
"""Count the triangles on the GPU.
The array toVertex is a flattened list of lists structure, where startEdge encodes the start indices of the lists.
:param startEdge: Indices into toVertex where edges start.
:type startEdge: list.
:param fromVertex: The from-vertex of each edge.
:type fromVertex: list.
:param toVertex: The to-vertex of each edge.
:type toVertex: list.
:return: Triangle count of graph.
:Examples:
.. doctest::
>>> from opveclib.examples.test_graph import load_graph_from_text_file, write_example_graph_to_text_file, triangles
>>> tmpName = "/tmp/v7e20.txt"
>>> write_example_graph_to_text_file(tmpName)
>>> startEdge, fromVertex, toVertex = load_graph_from_text_file(tmpName)
>>> triangles(startEdge, fromVertex, toVertex)
3
"""
if not ovl.local.cuda_enabled and target_language == 'cuda':
ovl.logger.debug("triangles tried to use cuda which is not enabled.")
return 0
count = ovl.evaluate(triangles_op(startEdge, fromVertex, toVertex), target_language=target_language)
return np.sum(count, axis=0, dtype=np.uint64)
def test(self):
"""
Test the correctness of ovl operator vs numpy implementation
"""
a = np.array([1e-10, -1e-10, 0.0, np.Infinity], dtype=np.float64)
expm1_op = expm1(a)
ref = np.expm1(a)
ovl_res = ovl.evaluate(expm1_op)
ovl.logger.info(u'numpy: ' + str(ref) + u' ovl: ' + str(ovl_res))
assert np.allclose(ref, ovl_res, rtol=0, atol=1e-20)
if ovl.cuda_enabled:
assert np.allclose(np.expm1(a),
ovl.evaluate(expm1_op, target_language='cuda'),
rtol=0,
atol=1e-20)
# test vs tensorflow
# ensure TF runs on GPU when asked
test_config = tf.ConfigProto(allow_soft_placement=False)
test_config.graph_options.optimizer_options.opt_level = -1
ones = np.ones_like(a)
if ovl.cuda_enabled:
devices = ['/cpu:0', '/gpu:0']
else:
devices = ['/cpu:0']
with tf.Session(config=test_config) as sess:
for dev_string in devices:
with tf.device(dev_string):
expm1_tf = ovl.as_tensorflow(expm1_op)
sess.run(tf.initialize_all_variables())
expm1_tf_result = sess.run(expm1_tf)
assert np.allclose(ref,
expm1_tf_result,
rtol=0,
atol=1e-20)
# TF exp - 1
tf_out = tf.exp(a) - ones
tf_result = tf_out.eval()
# this should fail
assert (np.allclose(ref, tf_result, rtol=0,
atol=1e-20) == False)
sess.close()
def test(self):
"""
Test the correctness of ovl operator vs numpy implementation
"""
a = np.array([1e-99, -1e-99, 0.0, np.Infinity], dtype=np.float64)
log1pOp = log1p(a)
ref = np.log1p(a)
ovl_res = ovl.evaluate(log1pOp)
ovl.logger.info(u'numpy: ' + str(ref) + u' ovl: ' + str(ovl_res))
assert np.allclose(ref, ovl_res, rtol=0, atol=1e-20)
if ovl.cuda_enabled:
assert np.allclose(np.log1p(a),
ovl.evaluate(log1pOp, target_language='cuda'),
rtol=0, atol=1e-20)
# test vs tensorflow
test_config=tf.ConfigProto(allow_soft_placement=False)
# ensure TF runs on GPU when asked
test_config.graph_options.optimizer_options.opt_level = -1
ones = np.ones_like(a)
if ovl.cuda_enabled:
devices = ['/cpu:0', '/gpu:0']
else:
devices = ['/cpu:0']
with tf.Session(config=test_config) as sess:
for dev_string in devices:
with tf.device(dev_string):
log1p_tf = ovl.as_tensorflow(log1pOp)
sess.run(tf.initialize_all_variables())
log1p_tf_result = sess.run(log1p_tf)
assert np.allclose(ref, log1p_tf_result,
rtol=0, atol=1e-20)
# TF exp - 1
tf_out = tf.log(a - ones)
tf_result = tf_out.eval()
# this should fail
assert (np.allclose(ref, tf_result,
rtol=0, atol=1e-20) == False)
sess.close()
def run_tests():
bb = 1
cc = 1
ee = 1000
k_num = 10
a1 = np.random.random((bb, ee, cc))
a2 = np.random.random((bb, ee, cc))
for k_ee in range(13, 14):
b = np.random.random((k_num, k_ee, cc))
for md in ['valid', 'same', 'full']:
for orientation in ['as-is', 'flipped']:
import time
t1 = time.time()
y1 = reference(a1, b, md, orientation, 'NEC')
t2 = time.time()
y2 = reference(a2, b, md, orientation, 'NEC')
op = conv_1d(a1,
b,
mode=md,
kernel_orientation=orientation,
data_format='NEC')
# result1 =
assert np.allclose(ovl.evaluate(op, target_language='cuda'),
y1)
# for d in range(1):
a1[:] = a2[:]
assert np.allclose(ovl.evaluate(op, target_language='cuda'),
y2)
res, prof = ovl.profile(op,
target_language='cuda',
profiling_iterations=100,
opt_level=3)
# print(prof)
# print(debug)
# print(op[0, 0, :])
ovl.logger.debug(k_ee, md, orientation, (t2 - t1) * 1000,
np.min(list(prof.values())[0]))
def test(self):
"""
Test the outputs of the operators to make sure they are consistent with the numpy implementation
"""
a = np.random.random((5, 5, 5))
ovl.logger.debug(u'Testing C')
assert np.allclose(
np.cumsum(a, axis=0),
ovl.evaluate(cumsum(a, axis=0), target_language='cpp'))
assert np.allclose(
np.cumsum(a, axis=1),
ovl.evaluate(cumsum(a, axis=1), target_language='cpp'))
assert np.allclose(
np.cumsum(a, axis=2),
ovl.evaluate(cumsum(a, axis=2), target_language='cpp'))
assert np.allclose(
np.cumprod(a, axis=0),
ovl.evaluate(cumprod(a, axis=0), target_language='cpp'))
assert np.allclose(
np.cumprod(a, axis=1),
ovl.evaluate(cumprod(a, axis=1), target_language='cpp'))
assert np.allclose(
np.cumprod(a, axis=2),
ovl.evaluate(cumprod(a, axis=2), target_language='cpp'))
if ovl.cuda_enabled:
ovl.logger.debug(u'Testing CUDA')
assert np.allclose(
np.cumsum(a, axis=0),
ovl.evaluate(cumsum(a, axis=0), target_language='cuda'))
assert np.allclose(
np.cumsum(a, axis=1),
ovl.evaluate(cumsum(a, axis=1), target_language='cuda'))
assert np.allclose(
np.cumsum(a, axis=2),
ovl.evaluate(cumsum(a, axis=2), target_language='cuda'))
assert np.allclose(
np.cumprod(a, axis=0),
ovl.evaluate(cumprod(a, axis=0), target_language='cuda'))
assert np.allclose(
np.cumprod(a, axis=1),
ovl.evaluate(cumprod(a, axis=1), target_language='cuda'))
assert np.allclose(
np.cumprod(a, axis=2),
ovl.evaluate(cumprod(a, axis=2), target_language='cuda'))
def test(self):
"""
Test the outputs of the operators to make sure they are consistent with the numpy implementation
"""
a = np.random.random((5, 5, 5))
ovl.logger.debug(u'Testing C')
assert np.allclose(np.cumsum(a, axis=0), ovl.evaluate(cumsum(a, axis=0), target_language='cpp'))
assert np.allclose(np.cumsum(a, axis=1), ovl.evaluate(cumsum(a, axis=1), target_language='cpp'))
assert np.allclose(np.cumsum(a, axis=2), ovl.evaluate(cumsum(a, axis=2), target_language='cpp'))
assert np.allclose(np.cumprod(a, axis=0), ovl.evaluate(cumprod(a, axis=0), target_language='cpp'))
assert np.allclose(np.cumprod(a, axis=1), ovl.evaluate(cumprod(a, axis=1), target_language='cpp'))
assert np.allclose(np.cumprod(a, axis=2), ovl.evaluate(cumprod(a, axis=2), target_language='cpp'))
if ovl.cuda_enabled:
ovl.logger.debug(u'Testing CUDA')
assert np.allclose(np.cumsum(a, axis=0), ovl.evaluate(cumsum(a, axis=0), target_language='cuda'))
assert np.allclose(np.cumsum(a, axis=1), ovl.evaluate(cumsum(a, axis=1), target_language='cuda'))
assert np.allclose(np.cumsum(a, axis=2), ovl.evaluate(cumsum(a, axis=2), target_language='cuda'))
assert np.allclose(np.cumprod(a, axis=0), ovl.evaluate(cumprod(a, axis=0), target_language='cuda'))
assert np.allclose(np.cumprod(a, axis=1), ovl.evaluate(cumprod(a, axis=1), target_language='cuda'))
assert np.allclose(np.cumprod(a, axis=2), ovl.evaluate(cumprod(a, axis=2), target_language='cuda'))