1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
|
from hypothesis import given
import hypothesis.strategies as st
import numpy as np
import unittest
from caffe2.python import core, workspace, dyndep
import caffe2.python.hypothesis_test_util as hu
dyndep.InitOpsLibrary("@/caffe2/caffe2/mpi:mpi_ops")
_has_mpi =False
COMM = None
RANK = 0
SIZE = 0
def SetupMPI():
try:
# pyre-fixme[21]: undefined import
from mpi4py import MPI
global _has_mpi, COMM, RANK, SIZE
_has_mpi = core.IsOperatorWithEngine("CreateCommonWorld", "MPI")
COMM = MPI.COMM_WORLD
RANK = COMM.Get_rank()
SIZE = COMM.Get_size()
except ImportError:
_has_mpi = False
@unittest.skipIf(not _has_mpi,
"MPI is not available. Skipping.")
class TestMPI(hu.HypothesisTestCase):
@given(X=hu.tensor(),
root=st.integers(min_value=0, max_value=SIZE - 1),
device_option=st.sampled_from(hu.device_options),
**hu.gcs)
def test_broadcast(self, X, root, device_option, gc, dc):
# Use mpi4py's broadcast to make sure that all nodes inherit the
# same hypothesis test.
X = COMM.bcast(X)
root = COMM.bcast(root)
device_option = COMM.bcast(device_option)
X[:] = RANK
self.assertTrue(
workspace.RunOperatorOnce(
core.CreateOperator(
"CreateCommonWorld", [], "comm", engine="MPI",
device_option=device_option)))
self.assertTrue(workspace.FeedBlob("X", X, device_option))
mpi_op = core.CreateOperator(
"Broadcast", ["comm", "X"], "X", engine="MPI", root=root,
device_option=device_option)
self.assertTrue(workspace.RunOperatorOnce(mpi_op))
new_X = workspace.FetchBlob("X")
np.testing.assert_array_equal(new_X, root)
workspace.ResetWorkspace()
@given(X=hu.tensor(),
root=st.integers(min_value=0, max_value=SIZE - 1),
device_option=st.sampled_from(hu.device_options),
**hu.gcs)
def test_reduce(self, X, root, device_option, gc, dc):
# Use mpi4py's broadcast to make sure that all nodes inherit the
# same hypothesis test.
X = COMM.bcast(X)
root = COMM.bcast(root)
device_option = COMM.bcast(device_option)
X[:] = RANK
self.assertTrue(
workspace.RunOperatorOnce(
core.CreateOperator(
"CreateCommonWorld", [], "comm", engine="MPI",
device_option=device_option)))
self.assertTrue(workspace.FeedBlob("X", X, device_option))
mpi_op = core.CreateOperator(
"Reduce", ["comm", "X"], "X_reduced", engine="MPI", root=root,
device_option=device_option)
self.assertTrue(workspace.RunOperatorOnce(mpi_op))
if (RANK == root):
new_X = workspace.FetchBlob("X")
np.testing.assert_array_equal(new_X, root)
workspace.ResetWorkspace()
@given(X=hu.tensor(),
root=st.integers(min_value=0, max_value=SIZE - 1),
device_option=st.sampled_from(hu.device_options),
inplace=st.booleans(),
**hu.gcs)
def test_allreduce(self, X, root, device_option, inplace, gc, dc):
# Use mpi4py's broadcast to make sure that all nodes inherit the
# same hypothesis test.
X = COMM.bcast(X)
root = COMM.bcast(root)
device_option = COMM.bcast(device_option)
inplace = COMM.bcast(inplace)
X[:] = RANK
self.assertTrue(
workspace.RunOperatorOnce(
core.CreateOperator(
"CreateCommonWorld", [], "comm", engine="MPI",
device_option=device_option)))
# Use mpi4py's broadcast to make sure that all copies have the same
# tensor size.
X = COMM.bcast(X)
X[:] = RANK
self.assertTrue(workspace.FeedBlob("X", X, device_option))
mpi_op = core.CreateOperator(
"Allreduce", ["comm", "X"],
"X" if inplace else "X_reduced",
engine="MPI", root=root,
device_option=device_option)
self.assertTrue(workspace.RunOperatorOnce(mpi_op))
new_X = workspace.FetchBlob("X" if inplace else "X_reduced")
np.testing.assert_array_equal(new_X, SIZE * (SIZE - 1) / 2)
workspace.ResetWorkspace()
@given(X=hu.tensor(),
device_option=st.sampled_from(hu.device_options),
specify_send_blob=st.booleans(),
specify_recv_blob=st.booleans(),
**hu.gcs)
def test_sendrecv(
self, X, device_option, specify_send_blob, specify_recv_blob,
gc, dc):
# Use mpi4py's broadcast to make sure that all nodes inherit the
# same hypothesis test.
X = COMM.bcast(X)
device_option = COMM.bcast(device_option)
specify_send_blob = COMM.bcast(specify_send_blob)
specify_recv_blob = COMM.bcast(specify_recv_blob)
X[:] = RANK
self.assertTrue(
workspace.RunOperatorOnce(
core.CreateOperator(
"CreateCommonWorld", [], "comm", engine="MPI",
device_option=device_option)))
self.assertTrue(workspace.FeedBlob("X", X, device_option))
for src in range(SIZE):
for dst in range(SIZE):
tag = src * SIZE + dst
if src == dst:
continue
elif RANK == src:
X[:] = RANK
self.assertTrue(workspace.FeedBlob("X", X, device_option))
if specify_send_blob:
self.assertTrue(workspace.FeedBlob(
"dst", np.array(dst, dtype=np.int32)))
self.assertTrue(workspace.FeedBlob(
"tag", np.array(tag, dtype=np.int32)))
mpi_op = core.CreateOperator(
"SendTensor", ["comm", "X", "dst", "tag"], [],
engine="MPI", raw_buffer=True,
device_option=device_option)
else:
mpi_op = core.CreateOperator(
"SendTensor", ["comm", "X"], [], engine="MPI",
dst=dst, tag=tag, raw_buffer=True,
device_option=device_option)
self.assertTrue(workspace.RunOperatorOnce(mpi_op))
elif RANK == dst:
if specify_recv_blob:
self.assertTrue(workspace.FeedBlob(
"src", np.array(src, dtype=np.int32)))
self.assertTrue(workspace.FeedBlob(
"tag", np.array(tag, dtype=np.int32)))
mpi_op = core.CreateOperator(
"ReceiveTensor", ["comm", "X", "src", "tag"],
["X", "src", "tag"],
engine="MPI",
src=src, tag=tag, raw_buffer=True,
device_option=device_option)
else:
mpi_op = core.CreateOperator(
"ReceiveTensor", ["comm", "X"], ["X", "src", "tag"],
engine="MPI",
src=src, tag=tag, raw_buffer=True,
device_option=device_option)
self.assertTrue(workspace.RunOperatorOnce(mpi_op))
received = workspace.FetchBlob("X")
np.testing.assert_array_equal(received, src)
src_blob = workspace.FetchBlob("src")
np.testing.assert_array_equal(src_blob, src)
tag_blob = workspace.FetchBlob("tag")
np.testing.assert_array_equal(tag_blob, tag)
# simply wait for the guys to finish
COMM.barrier()
workspace.ResetWorkspace()
if __name__ == "__main__":
SetupMPI()
import unittest
unittest.main()
|
