def testNotEnoughIpus(self):
def my_graph(pa, pb, pc):
with ipu.ops.ipu_shard(0):
o1 = pa + pb
with ipu.ops.ipu_shard(1):
o2 = pa + pc
with ipu.ops.ipu_shard(2):
out = o1 + o2
return out
with ops.device('cpu'):
pa = array_ops.placeholder(np.float32, [2], name="a")
pb = array_ops.placeholder(np.float32, [2], name="b")
pc = array_ops.placeholder(np.float32, [2], name="c")
report = gen_ipu_ops.ipu_event_trace()
with ops.device("/device:IPU:0"):
out = ipu_compiler.compile(my_graph, [pa, pb, pc])
cfg = ipu.utils.create_ipu_config(profiling=True)
cfg = ipu.utils.set_ipu_model_options(cfg, compile_ipu_code=False)
cfg = ipu.utils.auto_select_ipus(cfg, 2)
ipu.utils.configure_ipu_system(cfg)
with sl.Session() as sess:
with self.assertRaisesRegexp(errors.ResourceExhaustedError,
'Trying to compile a graph for'):
sess.run(out, {pa: [1., 1.], pb: [0., 1.], pc: [1., 5.]})
def testConvBackpropFilter(self):
with ops.device("/device:IPU:0"):
inp = array_ops.placeholder(np.float32, [2, 8, 8, 3])
fil = constant_op.constant([2, 2, 3, 5], np.int32)
bck = array_ops.placeholder(np.float32, [2, 8, 8, 5], name="wei")
output = nn_ops.conv2d_backprop_filter(
inp, fil, bck, strides=[1, 1, 1, 1], padding="SAME")
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
fd = {
inp: np.zeros([2, 8, 8, 3]),
bck: np.zeros([2, 8, 8, 5]),
}
result = sess.run(output, fd)
self.assertAllClose(result, np.zeros([2, 2, 3, 5]))
result = sess.run(report)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
ok = ['__seed*', 'Copy_', 'Conv2DBackpropFilter/convolution.*/Conv_8x8']
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testDepthwiseConvBackpropFilter1x1WithRelu(self):
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [1, 6, 6, 3], name="a")
pb = constant_op.constant([1, 1, 3, 2], dtype=np.int32) # filter sizes
pc = array_ops.placeholder(np.float32, [1, 6, 6, 6], name="c")
c = nn.depthwise_conv2d_native_backprop_filter(
pa, pb, pc, strides=[1, 1, 1, 1], padding="SAME")
c = nn.relu(c)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
fd = {pa: np.zeros([1, 6, 6, 3]), pc: np.zeros([1, 6, 6, 6])}
result = sess.run(c, fd)
self.assertAllClose(result, np.zeros([1, 1, 3, 2]))
result = sess.run(report)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
ok = [
'__seed*', 'Copy_',
'DepthwiseConv2dNativeBackpropFilter/fusion*/Conv_6x6',
'Relu/custom-call*/Nonlinearity'
]
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testIpuEventsWithoutPoplarReporting(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [2, 2], name="a")
pb = array_ops.placeholder(np.float32, [2, 2], name="b")
out = math_ops.add(pa, pb)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
opts = utils.create_ipu_config(profiling=False,
enable_ipu_events=True)
utils.configure_ipu_system(opts)
fd = {pa: [[1., 1.], [2., 3.]], pb: [[0., 1.], [4., 5.]]}
sess.run(report, fd)
sess.run(out, fd)
rep = sess.run(report, fd)
evts = utils.extract_all_events(rep)
self.assertEqual(len(evts),
3) # compile begin, compile end, execute
for e in evts:
if e.type == IpuTraceEvent.COMPILE_END:
self.assertFalse(e.compile_end.compilation_report)
if e.type == IpuTraceEvent.EXECUTE:
self.assertFalse(e.execute.execution_report)
sess.close()
def testDropoutImpl():
def ipu_dropout_back(w):
output = poprand.dropout(w, rate=0.4)
largest = output
cost = tf.square(largest)
opt = tf.train.GradientDescentOptimizer(learning_rate=0.1)
gradients = opt.compute_gradients(cost, w)
return [output, gradients]
with ops.device('cpu'):
input_data = array_ops.placeholder(np.float32, [32])
report = gen_ipu_ops.ipu_event_trace()
with ipu.ops.ipu_scope("/device:IPU:0"):
r = ipu_compiler.compile(ipu_dropout_back, inputs=[input_data])
cfg = ipu.utils.create_ipu_config()
cfg = ipu.utils.set_ipu_model_options(cfg, compile_ipu_code=False)
ipu.utils.configure_ipu_system(cfg)
with sl.Session() as sess:
in_data = np.random.rand(32)
out = sess.run(r, {input_data: in_data})
dropout_out = out[0]
gradients = out[1][0][0]
# Check we have the same number of zeros.
self.assertAllEqual(
np.count_nonzero(dropout_out), np.count_nonzero(gradients))
def testNamedOperations(self):
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [2, 2], name="a")
pb = array_ops.placeholder(np.float32, [2, 2], name="b")
with ops.name_scope('my_ops'):
out = math_ops.add(pa, pb, 'my_add_op')
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system()
with tu.ipu_session() as sess:
fd = {pa: [[1., 1.], [2., 3.]], pb: [[0., 1.], [4., 5.]]}
sess.run(report, fd)
result = sess.run(out, fd)
self.assertAllClose(result, [[1., 2.], [6., 8.]])
rep = sess.run(report, fd)
s = tu.extract_all_strings_from_event_trace(rep)
cs_list = tu.get_compute_sets_from_report(s)
ok = ['__seed*', 'my_ops/my_add_op/add']
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testReportEveryNthExecution_Every1(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [2, 2], name="a")
pb = array_ops.placeholder(np.float32, [2, 2], name="b")
out = math_ops.add(pa, pb)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
opts = utils.create_ipu_config(profiling=True,
profile_execution=True,
report_every_nth_execution=1,
use_poplar_text_report=False)
utils.configure_ipu_system(opts)
fd = {pa: [[1., 1.], [2., 3.]], pb: [[0., 1.], [4., 5.]]}
sess.run(report, fd)
sess.run(out, fd)
sess.run(out, fd)
sess.run(out, fd)
sess.run(out, fd)
sess.run(out, fd)
rep = sess.run(report, fd)
r = tu.ReportJSON(self)
types = r.parse_events(rep)
self.assertEqual(types[IpuTraceEvent.EXECUTE], 5)
self.assertEqual(len(r.get_execution_reports()), 5,
"Every execution should have generated a report")
def testBatchNormalizeLayerFusedFp16(self):
with ops.device("/device:IPU:0"):
with variable_scope.variable_scope("", use_resource=True):
x = array_ops.placeholder(np.float16, [4, 64, 64, 4], name="a")
normed = layers_norm.batch_normalization(x, fused=True)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
sess.run(variables.global_variables_initializer())
result = sess.run(normed, {x: np.zeros([4, 64, 64, 4])})
self.assertAllClose(result, np.zeros([4, 64, 64, 4]))
rep = sess.run(report)
s = tu.extract_all_strings_from_event_trace(rep)
cs = tu.get_compute_sets_from_report(s)
bl = ['*convert*/Cast*']
self.assertTrue(tu.check_compute_sets_not_in_blacklist(cs, bl))
def testMaxPool(self):
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [1, 1, 10, 10], name="a")
c = nn.max_pool(pa,
ksize=[1, 1, 5, 5],
strides=[1, 1, 2, 2],
data_format='NCHW',
padding='SAME',
name="max")
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
fd = {
pa: np.ones([1, 1, 10, 10]),
}
result = sess.run(c, fd)
self.assertAllClose(result, np.ones([1, 1, 5, 5]))
result = sess.run(report)
self.assertTrue(len(result) == 3)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
ok = ['__seed*', 'max/custom-call*/maxPool5x5']
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testArgMax(self):
batchsize = 4
n_categories = 1200
def model(a):
return math_ops.argmax(a, axis=1, output_type=dtypes.int32)
with ops.device('cpu'):
pa = array_ops.placeholder(np.float32, [batchsize, n_categories])
report = gen_ipu_ops.ipu_event_trace()
with ops.device("/device:IPU:0"):
out = model(pa)
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
input = np.random.rand(batchsize, n_categories)
fd = {pa: input}
result = sess.run(out, fd)
self.assertAllClose(result, np.argmax(input, axis=1))
result = sess.run(report)
self.assertTrue(len(result) == 3)
def testCheckMaxTileSize(self):
dtype = np.float32
shape = (1024, 2048)
with ops.device("/device:IPU:0"):
with variable_scope.variable_scope("", use_resource=True):
a = variable_scope.get_variable(
"a",
shape=shape,
initializer=init_ops.constant_initializer(2),
dtype=dtype)
pb = array_ops.placeholder(shape=shape, dtype=dtype, name="b")
c = constant_op.constant(4, shape=shape, dtype=dtype, name="c")
output = a + pb + c
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system(execution_trace=False)
with tu.ipu_session() as sess:
sess.run(variables.global_variables_initializer())
result = sess.run(report)
s = tu.extract_all_strings_from_event_trace(result)
max_tile_size = tu.get_maximum_tile_size_from_events(s)
self.assertTrue(max_tile_size < 17000)
out = sess.run(output, {pb: np.ones(shape=shape, dtype=dtype)})
self.assertAllClose(np.full(shape, 7, dtype=dtype), out)
result = sess.run(report)
s = tu.extract_all_strings_from_event_trace(result)
max_tile_size = tu.get_maximum_tile_size_from_events(s)
self.assertTrue(max_tile_size < 40000)
def testTopK(self):
n_categories = 1200
topn = 24
def model(a):
values, indices = nn.top_k(a, topn)
return indices
with ops.device('cpu'):
pa = array_ops.placeholder(np.float32, [n_categories], name="a")
report = gen_ipu_ops.ipu_event_trace()
with ops.device("/device:IPU:0"):
out = model(pa)
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
input = np.random.random(n_categories)
expected = (-input).argsort()[:topn]
fd = {pa: input}
result = sess.run(out, fd)
self.assertAllClose(result, expected)
result = sess.run(report)
self.assertTrue(len(result) == 3)
def testTraining(self):
x = array_ops.placeholder(datatype, shape=[1, 224, 224, 4])
y_ = array_ops.placeholder(datatype, shape=[1, 1000])
with ipu_ops.ipu_scope("/device:IPU:0"):
logits = inference(x)
loss = math_ops.reduce_mean(
nn_ops.softmax_cross_entropy_with_logits_v2(
logits=logits, labels=array_ops.stop_gradient(y_)))
train = gradient_descent.GradientDescentOptimizer(0.01).minimize(loss)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
opts = utils.create_ipu_config(profiling=True)
utils.configure_ipu_system(opts)
sess = sl.Session()
sess.run(variables.global_variables_initializer())
sess.run(report)
data = np.zeros([1, 224, 224, 4])
labels = np.zeros([1, 1000])
sess.run(train, feed_dict={x: data, y_: labels})
out = sess.run(report)
sess.close()
evts = utils.extract_all_events(out)
size = utils.get_memory_size_from_events(evts)
self.assertTrue(size < 174000000)
def testMultipleConfigureIpuShouldFail(self):
def my_graph(pa, pb, pc):
with ops.device("/device:IPU:0"):
o1 = pa + pb
o2 = pa + pc
out = o1 + o2
return [out]
with ops.device('cpu'):
pa = array_ops.placeholder(np.float32, [2], name="a")
pb = array_ops.placeholder(np.float32, [2], name="b")
pc = array_ops.placeholder(np.float32, [2], name="c")
report = gen_ipu_ops.ipu_event_trace()
out = ipu_compiler.compile(my_graph, [pa, pb, pc])
cfg = ipu.utils.create_ipu_config(profiling=True)
cfg = ipu.utils.set_ipu_model_options(cfg, compile_ipu_code=False)
cfg = ipu.utils.auto_select_ipus(cfg, 2)
ipu.utils.configure_ipu_system(cfg)
with self.assertRaises(Exception):
cfg = ipu.utils.create_ipu_config(profiling=True)
cfg = ipu.utils.set_ipu_model_options(cfg, compile_ipu_code=True)
ipu.utils.configure_ipu_system(cfg)
def testUniformRandomNonScalarInitalizer(self):
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
with ops.device("/device:IPU:0"):
with variable_scope.variable_scope("vs", use_resource=True):
i = init_ops.random_uniform_initializer(minval=-2.0,
maxval=2.0)
z = variable_scope.get_variable("z1",
shape=[2],
dtype=np.float32,
initializer=i)
tu.configure_ipu_system()
with tu.ipu_session() as sess:
# Clean existing reports
sess.run(report)
sess.run(variables.global_variables_initializer())
r = sess.run(report)
o = sess.run(z)
self.assertAllClose(o, [0.0, 0.0], 2.0, 2.0)
s = tu.extract_all_strings_from_event_trace(r)
cs_list = tu.get_compute_sets_from_report(s)
ok = [
'__seed*',
'vs/z1/Initializer/random_uniform/RandomUniform/fusion/uniform'
]
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testScaledSubtractFrom(self):
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float16, [3])
pb = array_ops.placeholder(np.float16, [3])
const = array_ops.constant(2.0, np.float16)
# note how const operand index varies compared to testScaledAddTo
# still should match as it will be reordered
c = pa - const * pb
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
fd = {pa: [2.0, 0.5, 1.0], pb: [1.0, 2.0, 3.0]}
result = sess.run(c, fd)
self.assertAllClose(result, [0.0, -3.5, -5.0])
result = sess.run(report)
self.assertTrue(len(result) == 3)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
ok = ['__seed*', 'host-exchange-local-copy-', 'sub/fusion/AddTo']
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testTuplesOfTuplesAreStreamed(self):
with ops.device("/device:IPU:0"):
with variable_scope.variable_scope("vs", use_resource=True):
pa = array_ops.placeholder(np.int64, [2, 2], name="a")
pb = array_ops.placeholder(np.int64, [2, 2], name="b")
pc = array_ops.placeholder(np.int64, [2, 2], name="c")
c = control_flow_ops.tuple((pa + pc, pb + pc))
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system(True, True, True)
with tu.ipu_session() as sess:
sess.run(report)
in0 = np.full((2, 2), 7)
in1 = np.full((2, 2), 6)
in2 = np.full((2, 2), 5)
fd = {
pa: in0,
pb: in1,
pc: in2,
}
out = sess.run(c, fd)
self.assertEqual(len(out), 2)
self.assertAllClose(out, (np.full((2, 2), 12), np.full(
(2, 2), 11)))
rep = sess.run(report)
io_evts = tu.extract_all_io_events(rep)
# No io_events implies the data was streamed
self.assertEqual(len(list(io_evts)), 0)
def testScaledSubtractFromVariable(self):
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float16, [3])
pb = array_ops.placeholder(np.float16, [3])
pc = array_ops.placeholder(np.float16, [1])
c = pa - pc * pb
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
fd = {pa: [2.0, 0.5, 1.0], pb: [1.0, 2.0, 3.0], pc: [2.0]}
result = sess.run(c, fd)
self.assertAllClose(result, [0.0, -3.5, -5.0])
result = sess.run(report)
self.assertTrue(len(result) == 3)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
ok = ['__seed*', 'host-exchange-local-copy-', 'sub/fusion/AddTo']
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testCborReport(self):
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [2, 2], name="a")
pb = array_ops.placeholder(np.float32, [2, 2], name="b")
out = math_ops.add(pa, pb)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system(text_report=False, cbor_report=True)
with tu.ipu_session() as sess:
fd = {pa: [[1., 1.], [2., 3.]], pb: [[0., 1.], [4., 5.]]}
sess.run(report, fd)
sess.run(out, fd)
rep = sess.run(report, fd)
evts = tu.extract_all_events(rep)
self.assertEqual(len(evts), 3) # begin, end, execute
self.assertEqual(evts[1].compile_end.compilation_report[0],
bytes(bytearray([217]))[0])
self.assertEqual(evts[2].execute.execution_report[0],
bytes(bytearray([217]))[0])
def testSigmoidNotInplace(self):
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [3], name="a")
c = math_ops.sigmoid(pa) + pa
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
fd = {pa: [-6.0, 0.0, 6.0]}
result = sess.run(c, fd)
self.assertAllClose(result, [-5.997527, 0.5, 6.997527])
result = sess.run(report)
self.assertTrue(len(result) == 3)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
ok = [
'__seed*', 'Sigmoid/custom-call/Nonlinearity',
'Copy_XLA_Args/arg0.*_to_Sigmoid/custom-call.clone/OnTileCopy-0',
'add/add.*/AddTo'
]
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testCborReport(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [2, 2], name="a")
pb = array_ops.placeholder(np.float32, [2, 2], name="b")
out = math_ops.add(pa, pb)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
opts = utils.create_ipu_config(profiling=True,
profile_execution=True,
use_poplar_text_report=False,
use_poplar_cbor_report=True)
utils.configure_ipu_system(opts)
fd = {pa: [[1., 1.], [2., 3.]], pb: [[0., 1.], [4., 5.]]}
sess.run(report, fd)
sess.run(out, fd)
rep = sess.run(report, fd)
evts = utils.extract_all_events(rep)
self.assertEqual(len(evts), 4) # engine, begin, end, execute
self.assertEqual(evts[1].compile_end.compilation_report[0],
bytes(bytearray([217]))[0])
self.assertEqual(evts[3].execute.execution_report[0],
bytes(bytearray([217]))[0])
def testSigmoidGrad(self):
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [3], name="grad")
pb = array_ops.placeholder(np.float32, [3], name="in")
c = gen_math_ops.sigmoid_grad(pa, pb)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
fd = {pa: [2.0, 0.5, 1.0], pb: [-1.0, 1.0, 6.0]}
result = sess.run(c, fd)
self.assertAllClose(result, [2.0, 0.25, 0.0])
result = sess.run(report)
self.assertTrue(len(result) == 3)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
ok = ['__seed*', 'SigmoidGrad/custom-call/NonLinearityGrad']
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testIpuModelDeviceWithMultipleReport(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [2, 2], name="a")
pb = array_ops.placeholder(np.float32, [2, 2], name="b")
out1 = pa + pb
out2 = pa - pb
with ops.device('cpu'):
with ops.control_dependencies([out1, out2]):
report = gen_ipu_ops.ipu_event_trace()
opts = utils.create_ipu_config(profiling=True,
profile_execution=True)
utils.configure_ipu_system(opts)
fd = {pa: [[1., 1.], [2., 3.]], pb: [[0., 1.], [4., 5.]]}
sess.run(report, fd)
result = sess.run(out1, fd)
self.assertAllClose(result, [[1., 2.], [6., 8.]])
result, rep = sess.run([out2, report], fd)
self.assertAllClose(result, [[1., 0.], [-2., -2.]])
# 2x engine, 2x compile_begin, 2x compile_end, 2x load engine
self.assertEqual(len(rep), 8)
def testDontOutlineInplaceExpression(self):
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [])
pb = array_ops.placeholder(np.float32, [])
pc = array_ops.placeholder(np.float32, [])
pd = array_ops.placeholder(np.float32, [])
e = pa + pb - pc + pd
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
fd = {pa: 1, pb: 2, pc: 3, pd: 4}
result = sess.run(e, fd)
self.assertAllClose(result, 4)
result = sess.run(report)
self.assertTrue(len(result) == 3)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
ok = [
'__seed*', 'add/add.*/AddTo', 'sub/subtract.*/AddTo',
'add_1/add.*/AddTo'
]
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testDropoutImpl(rate):
def ipu_dropout(w):
output = poprand.dropout(w, rate=rate)
return [output]
with ops.device('cpu'):
input_data = array_ops.placeholder(np.float32, [1024, 1024, 4])
report = gen_ipu_ops.ipu_event_trace()
with ipu.ops.ipu_scope("/device:IPU:0"):
r = ipu_compiler.compile(ipu_dropout, inputs=[input_data])
cfg = ipu.utils.create_ipu_config()
cfg = ipu.utils.set_ipu_model_options(cfg, compile_ipu_code=False)
ipu.utils.configure_ipu_system(cfg)
with sl.Session() as sess:
in_data = np.random.rand(1024, 1024, 4)
result = sess.run(r, {input_data: in_data})
percent_kept = np.count_nonzero(result) / np.count_nonzero(in_data)
# There's a considerable amount for randomness so we have a reasonably large
# dimensionality of test data to make sure the error is smaller.
is_roughly_close = abs(percent_kept - (1.0 - rate))
# The observed error is actually a lot less than this (>1%) but we don't want to cause
# random regressions and 3% is probably still acceptable for any outlier randoms.
self.assertTrue(is_roughly_close < 0.03)
def testRandomNormalInitalizer(self):
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
with ops.device("/device:IPU:0"):
with variable_scope.variable_scope("vs", use_resource=True):
i = init_ops.random_normal_initializer(mean=2.0, stddev=0.01)
z = variable_scope.get_variable("z1",
shape=[],
dtype=np.float32,
initializer=i)
tu.configure_ipu_system()
with tu.ipu_session() as sess:
# Clean existing reports
sess.run(report)
sess.run(variables.global_variables_initializer())
r = sess.run(report)
o = sess.run(z)
self.assertAllClose(o, 2.0, 0.2, 0.2)
s = tu.extract_all_strings_from_event_trace(r)
cs_list = tu.get_compute_sets_from_report(s)
ok = [
'__seed*',
'vs/z1/Initializer/random_normal/RandomStandardNormal/fusion/normal'
]
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testDepthwiseConvBackpropInput1x1(self):
with ops.device("/device:IPU:0"):
pa = constant_op.constant([1, 8, 8, 3], dtype=np.int32) # input sizes
pb = array_ops.placeholder(np.float32, [1, 1, 3, 2], name="b")
pc = array_ops.placeholder(np.float32, [1, 8, 8, 6], name="c")
c = nn.depthwise_conv2d_native_backprop_input(
pa, pb, pc, strides=[1, 1, 1, 1], padding="SAME")
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(report)
fd = {pb: np.zeros([1, 1, 3, 2]), pc: np.zeros([1, 8, 8, 6])}
result = sess.run(c, fd)
self.assertAllClose(result, np.zeros([1, 8, 8, 3]))
result = sess.run(report)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
ok = [
'__seed*',
'DepthwiseConv2dNativeBackpropInput/fusion*/WeightTranspose',
'DepthwiseConv2dNativeBackpropInput/fusion*/Conv_1x1', 'Copy_'
]
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testDefaultTruncatedNormalInitalizer(self):
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
with ops.device("/device:IPU:0"):
with variable_scope.variable_scope("", use_resource=True):
i = init_ops.truncated_normal_initializer()
z = variable_scope.get_variable("z1",
shape=[2, 4],
dtype=np.float32,
initializer=i)
tu.configure_ipu_system()
with tu.ipu_session() as sess:
sess.run(variables.global_variables_initializer())
o = sess.run(z)
self.assertAllClose(o, np.ones((2, 4)), 2.0, 2.0)
# Find of the names of compute sets
r = sess.run(report)
s = tu.extract_all_strings_from_event_trace(r)
cs_list = tu.get_compute_sets_from_report(s)
ok = [
'__seed*',
'z1/Initializer/truncated_normal/TruncatedNormal/custom-call*/truncatedNormal'
]
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def ipu_compile_summary(name, op_list, collections=None):
"""Create an IPU compiler summary operation.
Args:
name: A name for the summary.
op_list: An operation or list of operations to make this summary dependent
upon.
collections: Optional collections to add the summary into.
Returns:
The new summary operation
"""
if not isinstance(op_list, list):
op_list = [op_list]
with ops.device("cpu"):
with ops.control_dependencies(op_list):
reports = gen_ipu_ops.ipu_event_trace()
summary_metadata = summary_pb2.SummaryMetadata(
plugin_data=summary_pb2.SummaryMetadata.PluginData(
plugin_name="ipu"))
t_summary = tensor_summary(name='ipu_trace',
tensor=reports,
summary_metadata=summary_metadata,
collections=collections,
display_name=name)
return t_summary
def testMultiScopeTest(self):
with ops.device('cpu'):
x = array_ops.placeholder(np.float32, [2, 2])
y = array_ops.placeholder(np.float32, [2, 2])
report = gen_ipu_ops.ipu_event_trace()
with ipu.scopes.ipu_scope('/device:IPU:0'):
z = math_ops.matmul(x, y)
with ipu.scopes.ipu_scope('/device:IPU:0'):
z2 = math_ops.matmul(x, z)
cfg = ipu.utils.create_ipu_config(profiling=True)
cfg = ipu.utils.set_ipu_model_options(cfg, compile_ipu_code=False)
ipu.utils.configure_ipu_system(cfg)
with sl.Session() as sess:
sess.run(report)
result = sess.run(z2, {x: np.ones([2, 2]), y: np.ones([2, 2])})
self.assertAllEqual(result, [[4, 4], [4, 4]])
rep = sess.run(report)
evts = ipu.utils.extract_all_types_from_event_trace(rep)
num_compiles = 0
num_executions = 0
for e in evts:
if e == IpuTraceEvent.COMPILE_END:
num_compiles += 1
if e == IpuTraceEvent.EXECUTE:
num_executions += 1
self.assertEqual(num_compiles, 1)
self.assertEqual(num_executions, 1)