forked from horovod/horovod
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsetup.py
1613 lines (1365 loc) · 65.9 KB
/
setup.py
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
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# Copyright 2019 Uber Technologies, Inc. All Rights Reserved.
# Modifications copyright Microsoft
# Modifications copyright (C) 2020, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
import os
import re
import shlex
import subprocess
import sys
import textwrap
import traceback
import pipes
import warnings
from copy import deepcopy
from distutils.errors import CompileError, DistutilsError, \
DistutilsPlatformError, LinkError
from distutils.sysconfig import customize_compiler
from distutils.version import LooseVersion
from setuptools import setup, Extension, find_packages
from setuptools.command.build_ext import build_ext
from horovod import __version__
from horovod.common.util import env
class CMakeExtension(Extension):
def __init__(self, name, cmake_lists_dir='.', sources=[], **kwa):
Extension.__init__(self, name, sources=sources, **kwa)
self.cmake_lists_dir = os.path.abspath(cmake_lists_dir)
tensorflow_mpi_lib = Extension('horovod.tensorflow.mpi_lib', [])
torch_mpi_lib = Extension('horovod.torch.mpi_lib', [])
torch_mpi_lib_impl = Extension('horovod.torch.mpi_lib_impl', [])
torch_mpi_lib_v2 = Extension('horovod.torch.mpi_lib_v2', [])
mxnet_mpi_lib = Extension('horovod.mxnet.mpi_lib', [])
gloo_lib = CMakeExtension('gloo', cmake_lists_dir='third_party/gloo',
sources=[])
horovod_cuda_lib = CMakeExtension('horovod_cuda_kernels',
cmake_lists_dir='horovod/common/ops/cuda',
sources=[])
ccl_root = os.environ.get('CCL_ROOT')
have_ccl = ccl_root is not None
def is_build_action():
if len(sys.argv) <= 1:
return False
if sys.argv[1].startswith('build'):
return True
if sys.argv[1].startswith('bdist'):
return True
if sys.argv[1].startswith('install'):
return True
def check_tf_version():
try:
import tensorflow as tf
if LooseVersion(tf.__version__) < LooseVersion('1.15.0'):
raise DistutilsPlatformError(
'Your TensorFlow version %s is outdated. '
'Horovod requires tensorflow>=1.15.0' % tf.__version__)
# parse version
version = parse_version(tf.__version__)
if version is None:
raise DistutilsPlatformError(
'Unable to determine TensorFlow version from the version string \'%s\'' % tf.__version__)
return version
except ImportError:
raise DistutilsPlatformError(
'import tensorflow failed, is it installed?\n\n%s' % traceback.format_exc())
except AttributeError:
# This means that tf.__version__ was not exposed, which makes it *REALLY* old.
raise DistutilsPlatformError(
'Your TensorFlow version is outdated. Horovod requires tensorflow>=1.15.0')
def check_mx_version():
try:
import mxnet as mx
if mx.__version__ < '1.4.0':
raise DistutilsPlatformError(
'Your MXNet version %s is outdated. '
'Horovod requires mxnet>=1.4.0' % mx.__version__)
except ImportError:
raise DistutilsPlatformError(
'import mxnet failed, is it installed?\n\n%s' % traceback.format_exc())
except AttributeError:
raise DistutilsPlatformError(
'Your MXNet version is outdated. Horovod requires mxnet>1.3.0')
def get_supported_instruction_set_flags(flags_to_check):
supported = []
try:
flags_output = subprocess.check_output(
'gcc -march=native -E -v - </dev/null 2>&1 | grep cc1',
shell=True, universal_newlines=True).strip()
flags = shlex.split(flags_output)
supported = [x for x in flags_to_check if x in flags or x.replace('-m', '+') in flags]
except subprocess.CalledProcessError:
# Fallback to no advanced instruction set flags if were not able to get flag information.
pass
return supported
def get_cpp_flags(build_ext):
last_err = None
default_flags = ['-std=c++11', '-fPIC', '-O3', '-Wall', '-fassociative-math', '-ffast-math', '-ftree-vectorize', '-funsafe-math-optimizations']
build_arch_flags_env = os.environ.get('HOROVOD_BUILD_ARCH_FLAGS')
build_arch_flags = get_supported_instruction_set_flags(['-mf16c', '-mavx', '-mfma']) if build_arch_flags_env is None else build_arch_flags_env.split()
if sys.platform == 'darwin':
# Darwin most likely will have Clang, which has libc++.
flags_to_try = [default_flags + ['-stdlib=libc++'] + build_arch_flags,
default_flags + build_arch_flags,
default_flags + ['-stdlib=libc++'],
default_flags]
else:
flags_to_try = [default_flags + build_arch_flags,
default_flags + ['-stdlib=libc++'] + build_arch_flags,
default_flags,
default_flags + ['-stdlib=libc++']]
for cpp_flags in flags_to_try:
try:
test_compile(build_ext, 'test_cpp_flags',
extra_compile_preargs=cpp_flags,
code=textwrap.dedent('''\
#include <unordered_map>
void test() {
}
'''))
return cpp_flags
except (CompileError, LinkError):
last_err = 'Unable to determine C++ compilation flags (see error above).'
except Exception:
last_err = 'Unable to determine C++ compilation flags. ' \
'Last error:\n\n%s' % traceback.format_exc()
raise DistutilsPlatformError(last_err)
def get_nvcc_flags():
default_flags = ['--std=c++11', '-O3', '-Xcompiler', '-fPIC']
cc_list_env = os.environ.get('HOROVOD_BUILD_CUDA_CC_LIST')
# Invoke nvcc and extract all supported compute capabilities for CUDA toolkit version
full_cc_list = subprocess.check_output("nvcc --help | sed -n -e '/gpu-architecture <arch>/,/gpu-code <code>/ p' | sed -n -e '/Allowed values/,/gpu-code <code>/ p' | "
"grep -i sm_ | grep -Eo 'sm_[0-9]+' | sed -e s/sm_//g | sort -g -u | tr '\n' ' '",
shell=True).strip().split()
full_cc_list = [int(i) for i in full_cc_list]
# Build native kernels for specified compute capabilities
cc_list = full_cc_list if cc_list_env is None else [int(x) for x in cc_list_env.split(',')]
for cc in cc_list:
default_flags += ['-gencode', 'arch=compute_{cc},code=sm_{cc}'.format(cc=cc)]
# Build PTX for maximum specified compute capability
default_flags += ['-gencode', 'arch=compute_{cc},code=compute_{cc}'.format(cc=max(cc_list))]
return default_flags
def get_link_flags(build_ext):
last_err = None
libtool_flags = ['-Wl,-exported_symbols_list,horovod.exp']
ld_flags = ['-Wl,--version-script=horovod.lds']
if sys.platform == 'darwin':
flags_to_try = [libtool_flags, ld_flags]
else:
flags_to_try = [ld_flags, libtool_flags]
for link_flags in flags_to_try:
try:
test_compile(build_ext, 'test_link_flags',
extra_link_preargs=link_flags,
code=textwrap.dedent('''\
void test() {
}
'''))
return link_flags
except (CompileError, LinkError):
last_err = 'Unable to determine C++ link flags (see error above).'
except Exception:
last_err = 'Unable to determine C++ link flags. ' \
'Last error:\n\n%s' % traceback.format_exc()
raise DistutilsPlatformError(last_err)
def get_tf_include_dirs():
import tensorflow as tf
tf_inc = tf.sysconfig.get_include()
return [tf_inc, '%s/external/nsync/public' % tf_inc]
def get_tf_lib_dirs():
import tensorflow as tf
tf_lib = tf.sysconfig.get_lib()
return [tf_lib]
def get_tf_libs(build_ext, lib_dirs, cpp_flags):
last_err = None
for tf_libs in [['tensorflow_framework'], []]:
try:
lib_file = test_compile(build_ext, 'test_tensorflow_libs',
library_dirs=lib_dirs, libraries=tf_libs,
extra_compile_preargs=cpp_flags,
code=textwrap.dedent('''\
void test() {
}
'''))
from tensorflow.python.framework import load_library
load_library.load_op_library(lib_file)
return tf_libs
except (CompileError, LinkError):
last_err = 'Unable to determine -l link flags to use with TensorFlow (see error above).'
except Exception:
last_err = 'Unable to determine -l link flags to use with TensorFlow. ' \
'Last error:\n\n%s' % traceback.format_exc()
raise DistutilsPlatformError(last_err)
def get_tf_abi(build_ext, include_dirs, lib_dirs, libs, cpp_flags):
last_err = None
cxx11_abi_macro = '_GLIBCXX_USE_CXX11_ABI'
for cxx11_abi in ['0', '1']:
try:
lib_file = test_compile(build_ext, 'test_tensorflow_abi',
macros=[(cxx11_abi_macro, cxx11_abi)],
include_dirs=include_dirs,
library_dirs=lib_dirs,
libraries=libs,
extra_compile_preargs=cpp_flags,
code=textwrap.dedent('''\
#include <string>
#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/shape_inference.h"
void test() {
auto ignore = tensorflow::strings::StrCat("a", "b");
}
'''))
from tensorflow.python.framework import load_library
load_library.load_op_library(lib_file)
return cxx11_abi_macro, cxx11_abi
except (CompileError, LinkError):
last_err = 'Unable to determine CXX11 ABI to use with TensorFlow (see error above).'
except Exception:
last_err = 'Unable to determine CXX11 ABI to use with TensorFlow. ' \
'Last error:\n\n%s' % traceback.format_exc()
raise DistutilsPlatformError(last_err)
def get_tf_flags(build_ext, cpp_flags):
import tensorflow as tf
try:
return tf.sysconfig.get_compile_flags(), tf.sysconfig.get_link_flags()
except AttributeError:
# fallback to the previous logic
tf_include_dirs = get_tf_include_dirs()
tf_lib_dirs = get_tf_lib_dirs()
tf_libs = get_tf_libs(build_ext, tf_lib_dirs, cpp_flags)
tf_abi = get_tf_abi(build_ext, tf_include_dirs,
tf_lib_dirs, tf_libs, cpp_flags)
compile_flags = []
for include_dir in tf_include_dirs:
compile_flags.append('-I%s' % include_dir)
if tf_abi:
compile_flags.append('-D%s=%s' % tf_abi)
link_flags = []
for lib_dir in tf_lib_dirs:
link_flags.append('-L%s' % lib_dir)
for lib in tf_libs:
link_flags.append('-l%s' % lib)
return compile_flags, link_flags
def get_mx_include_dirs():
import mxnet as mx
return [mx.libinfo.find_include_path()]
def get_mx_lib_dirs():
import mxnet as mx
mx_libs = mx.libinfo.find_lib_path()
mx_lib_dirs = [os.path.dirname(mx_lib) for mx_lib in mx_libs]
return mx_lib_dirs
def get_mx_libs(build_ext, lib_dirs, cpp_flags):
last_err = None
for mx_libs in [['mxnet'], []]:
try:
lib_file = test_compile(build_ext, 'test_mx_libs',
library_dirs=lib_dirs, libraries=mx_libs,
extra_compile_preargs=cpp_flags,
code=textwrap.dedent('''\
void test() {
}
'''))
return mx_libs
except (CompileError, LinkError):
last_err = 'Unable to determine -l link flags to use with MXNet (see error above).'
except Exception:
last_err = 'Unable to determine -l link flags to use with MXNet. ' \
'Last error:\n\n%s' % traceback.format_exc()
raise DistutilsPlatformError(last_err)
def get_mx_flags(build_ext, cpp_flags):
mx_include_dirs = get_mx_include_dirs()
mx_lib_dirs = get_mx_lib_dirs()
mx_libs = get_mx_libs(build_ext, mx_lib_dirs, cpp_flags)
compile_flags = []
has_mkldnn = is_mx_mkldnn()
for include_dir in mx_include_dirs:
compile_flags.append('-I%s' % include_dir)
if has_mkldnn:
mkldnn_include = os.path.join(include_dir, 'mkldnn')
compile_flags.append('-I%s' % mkldnn_include)
link_flags = []
for lib_dir in mx_lib_dirs:
link_flags.append('-Wl,-rpath,%s' % lib_dir)
link_flags.append('-L%s' % lib_dir)
for lib in mx_libs:
link_flags.append('-l%s' % lib)
return compile_flags, link_flags
def get_mpi_flags():
show_command = os.environ.get('HOROVOD_MPICXX_SHOW', 'mpicxx -show')
try:
mpi_show_output = subprocess.check_output(
shlex.split(show_command), universal_newlines=True).strip()
mpi_show_args = shlex.split(mpi_show_output)
if not mpi_show_args[0].startswith('-'):
# Open MPI and MPICH print compiler name as a first word, skip it
mpi_show_args = mpi_show_args[1:]
# strip off compiler call portion and always escape each arg
return ' '.join(['"' + arg.replace('"', '"\'"\'"') + '"'
for arg in mpi_show_args])
except Exception:
raise DistutilsPlatformError(
'%s failed (see error below), is MPI in $PATH?\n'
'Note: If your version of MPI has a custom command to show compilation flags, '
'please specify it with the HOROVOD_MPICXX_SHOW environment variable.\n\n'
'%s' % (show_command, traceback.format_exc()))
def test_compile(build_ext, name, code, libraries=None, include_dirs=None,
library_dirs=None,
macros=None, extra_compile_preargs=None,
extra_link_preargs=None):
test_compile_dir = os.path.join(build_ext.build_temp, 'test_compile')
if not os.path.exists(test_compile_dir):
os.makedirs(test_compile_dir)
source_file = os.path.join(test_compile_dir, '%s.cc' % name)
with open(source_file, 'w') as f:
f.write(code)
compiler = build_ext.compiler
[object_file] = compiler.object_filenames([source_file])
shared_object_file = compiler.shared_object_filename(
name, output_dir=test_compile_dir)
compiler.compile([source_file], extra_preargs=extra_compile_preargs,
include_dirs=include_dirs, macros=macros)
compiler.link_shared_object(
[object_file], shared_object_file, libraries=libraries,
library_dirs=library_dirs,
extra_preargs=extra_link_preargs)
return shared_object_file
def get_cuda_dirs(build_ext, cpp_flags):
cuda_include_dirs = []
cuda_lib_dirs = []
cuda_home = os.environ.get('HOROVOD_CUDA_HOME')
if cuda_home:
cuda_include_dirs += ['%s/include' % cuda_home]
cuda_lib_dirs += ['%s/lib' % cuda_home, '%s/lib64' % cuda_home]
cuda_include = os.environ.get('HOROVOD_CUDA_INCLUDE')
if cuda_include:
cuda_include_dirs += [cuda_include]
cuda_lib = os.environ.get('HOROVOD_CUDA_LIB')
if cuda_lib:
cuda_lib_dirs += [cuda_lib]
if not cuda_include_dirs and not cuda_lib_dirs:
# default to /usr/local/cuda
cuda_include_dirs += ['/usr/local/cuda/include']
cuda_lib_dirs += ['/usr/local/cuda/lib', '/usr/local/cuda/lib64']
try:
test_compile(build_ext, 'test_cuda', libraries=['cudart'],
include_dirs=cuda_include_dirs,
library_dirs=cuda_lib_dirs,
extra_compile_preargs=cpp_flags,
code=textwrap.dedent('''\
#include <cuda_runtime.h>
void test() {
cudaSetDevice(0);
}
'''))
except (CompileError, LinkError):
raise DistutilsPlatformError(
'CUDA library was not found (see error above).\n'
'Please specify correct CUDA location with the HOROVOD_CUDA_HOME '
'environment variable or combination of HOROVOD_CUDA_INCLUDE and '
'HOROVOD_CUDA_LIB environment variables.\n\n'
'HOROVOD_CUDA_HOME - path where CUDA include and lib directories can be found\n'
'HOROVOD_CUDA_INCLUDE - path to CUDA include directory\n'
'HOROVOD_CUDA_LIB - path to CUDA lib directory')
return cuda_include_dirs, cuda_lib_dirs
def get_rocm_dirs(build_ext, cpp_flags):
rocm_include_dirs = []
rocm_lib_dirs = []
rocm_libs = ['hip_hcc']
rocm_macros = [('__HIP_PLATFORM_HCC__',1)]
rocm_path = os.environ.get('HOROVOD_ROCM_HOME', '/opt/rocm')
rocm_include_dirs += [
'%s/include' % rocm_path,
'%s/hcc/include' % rocm_path,
'%s/hip/include' % rocm_path,
'%s/hsa/include' % rocm_path,
]
rocm_lib_dirs += [
'%s/lib' % rocm_path,
]
try:
test_compile(build_ext, 'test_hip', libraries=rocm_libs, include_dirs=rocm_include_dirs,
library_dirs=rocm_lib_dirs, extra_compile_preargs=cpp_flags, macros=rocm_macros,
code=textwrap.dedent('''\
#include <hip/hip_runtime.h>
void test() {
hipSetDevice(0);
}
'''))
except (CompileError, LinkError):
raise DistutilsPlatformError(
'HIP library and/or ROCm header files not found (see error above).\n'
'Please specify correct ROCm location with the HOROVOD_ROCM_HOME environment variable')
return rocm_include_dirs, rocm_lib_dirs, rocm_macros
def get_nccl_vals(build_ext, gpu_include_dirs, gpu_lib_dirs, gpu_macros, cpp_flags, have_rocm):
nccl_include_dirs = []
nccl_lib_dirs = []
nccl_libs = []
nccl_home = os.environ.get('HOROVOD_NCCL_HOME')
if nccl_home:
nccl_include_dirs += ['%s/include' % nccl_home]
nccl_lib_dirs += ['%s/lib' % nccl_home, '%s/lib64' % nccl_home]
nccl_include_dir = os.environ.get('HOROVOD_NCCL_INCLUDE')
if nccl_include_dir:
nccl_include_dirs += [nccl_include_dir]
nccl_lib_dir = os.environ.get('HOROVOD_NCCL_LIB')
if nccl_lib_dir:
nccl_lib_dirs += [nccl_lib_dir]
nccl_link_mode = os.environ.get('HOROVOD_NCCL_LINK', 'SHARED' if have_rocm else 'STATIC')
if nccl_link_mode.upper() == 'SHARED':
if have_rocm:
nccl_libs += ['rccl']
else:
nccl_libs += ['nccl']
else:
nccl_libs += ['nccl_static']
if have_rocm:
raise DistutilsPlatformError('RCCL must be a shared library')
try:
test_compile(build_ext, 'test_nccl', libraries=nccl_libs,
include_dirs=nccl_include_dirs + gpu_include_dirs,
library_dirs=nccl_lib_dirs + gpu_lib_dirs,
extra_compile_preargs=cpp_flags,
macros=gpu_macros,
code=textwrap.dedent('''\
#include <%s>
#if NCCL_MAJOR < 2
#error Horovod requires NCCL 2.0 or later version, please upgrade.
#endif
void test() {
ncclUniqueId nccl_id;
ncclGetUniqueId(&nccl_id);
}
'''%('rccl.h' if have_rocm else 'nccl.h')))
except (CompileError, LinkError):
raise DistutilsPlatformError(
'NCCL 2.0 library or its later version was not found (see error above).\n'
'Please specify correct NCCL location with the HOROVOD_NCCL_HOME '
'environment variable or combination of HOROVOD_NCCL_INCLUDE and '
'HOROVOD_NCCL_LIB environment variables.\n\n'
'HOROVOD_NCCL_HOME - path where NCCL include and lib directories can be found\n'
'HOROVOD_NCCL_INCLUDE - path to NCCL include directory\n'
'HOROVOD_NCCL_LIB - path to NCCL lib directory')
return nccl_include_dirs, nccl_lib_dirs, nccl_libs
def get_ddl_dirs(build_ext, cuda_include_dirs, cuda_lib_dirs, cpp_flags):
ddl_include_dirs = []
ddl_lib_dirs = []
ddl_home = os.environ.get('HOROVOD_DDL_HOME')
if ddl_home:
ddl_include_dirs += ['%s/include' % ddl_home]
ddl_lib_dirs += ['%s/lib' % ddl_home, '%s/lib64' % ddl_home]
ddl_include_dir = os.environ.get('HOROVOD_DDL_INCLUDE')
if ddl_include_dir:
ddl_include_dirs += [ddl_include_dir]
ddl_lib_dir = os.environ.get('HOROVOD_DDL_LIB')
if ddl_lib_dir:
ddl_lib_dirs += [ddl_lib_dir]
# Keep DDL legacy folders for backward compatibility
if not ddl_include_dirs:
ddl_include_dirs += ['/opt/DL/ddl/include']
if not ddl_lib_dirs:
ddl_lib_dirs += ['/opt/DL/ddl/lib']
try:
test_compile(build_ext, 'test_ddl', libraries=['ddl', 'ddl_pack'],
include_dirs=ddl_include_dirs + cuda_include_dirs,
library_dirs=ddl_lib_dirs + cuda_lib_dirs,
extra_compile_preargs=cpp_flags,
code=textwrap.dedent('''\
#include <ddl.hpp>
void test() {
}
'''))
except (CompileError, LinkError):
raise DistutilsPlatformError(
'IBM PowerAI DDL library was not found (see error above).\n'
'Please specify correct DDL location with the HOROVOD_DDL_HOME '
'environment variable or combination of HOROVOD_DDL_INCLUDE and '
'HOROVOD_DDL_LIB environment variables.\n\n'
'HOROVOD_DDL_HOME - path where DDL include and lib directories can be found\n'
'HOROVOD_DDL_INCLUDE - path to DDL include directory\n'
'HOROVOD_DDL_LIB - path to DDL lib directory')
return ddl_include_dirs, ddl_lib_dirs
def set_cuda_options(build_ext, COMPILE_FLAGS, MACROS, INCLUDES, SOURCES, BUILD_MPI, LIBRARY_DIRS, LIBRARIES,
NVCC_COMPILE_FLAGS,**kwargs):
cuda_include_dirs, cuda_lib_dirs = get_cuda_dirs(build_ext, COMPILE_FLAGS)
MACROS += [('HAVE_CUDA', '1'), ('HAVE_GPU', '1')]
INCLUDES += cuda_include_dirs
SOURCES += ['horovod/common/ops/cuda_operations.cc',
'horovod/common/ops/gpu_operations.cc']
if BUILD_MPI:
SOURCES += ['horovod/common/ops/mpi_gpu_operations.cc']
LIBRARY_DIRS += cuda_lib_dirs
LIBRARIES += ['cudart']
NVCC_COMPILE_FLAGS += get_nvcc_flags()
def get_common_options(build_ext):
cpp_flags = get_cpp_flags(build_ext)
link_flags = get_link_flags(build_ext)
is_mac = sys.platform == 'darwin'
compile_without_gloo = os.environ.get('HOROVOD_WITHOUT_GLOO')
if compile_without_gloo:
print('INFO: HOROVOD_WITHOUT_GLOO detected, skip compiling Horovod with Gloo.')
have_gloo = False
have_cmake = False
else:
# determining if system has cmake installed
compile_with_gloo = os.environ.get('HOROVOD_WITH_GLOO')
try:
cmake_bin = get_cmake_bin()
subprocess.check_output([cmake_bin, '--version'])
have_cmake = True
except Exception:
if compile_with_gloo:
# Require Gloo to succeed, otherwise fail the install.
raise RuntimeError('Cannot find CMake. CMake is required to build Horovod with Gloo.')
print('INFO: Cannot find CMake, will skip compiling Horovod with Gloo.')
have_cmake = False
# TODO: remove system check if gloo support MacOX in the future
# https://github.com/facebookincubator/gloo/issues/182
if is_mac:
if compile_with_gloo:
raise RuntimeError('Gloo cannot be compiled on MacOS. Unset HOROVOD_WITH_GLOO to use MPI.')
print('INFO: Gloo cannot be compiled on MacOS, will skip compiling Horovod with Gloo.')
have_gloo = not is_mac and have_cmake
compile_without_mpi = os.environ.get('HOROVOD_WITHOUT_MPI')
mpi_flags = ''
if compile_without_mpi:
print('INFO: HOROVOD_WITHOUT_MPI detected, skip compiling Horovod with MPI.')
have_mpi = False
else:
# If without_mpi flag is not set by user, try to get mpi flags
try:
mpi_flags = get_mpi_flags()
have_mpi = True
except Exception:
if os.environ.get('HOROVOD_WITH_MPI'):
# Require MPI to succeed, otherwise fail the install.
raise
# If exceptions happen, will not use mpi during compilation.
print(traceback.format_exc(), file=sys.stderr)
print('INFO: Cannot find MPI compilation flags, will skip compiling with MPI.')
have_mpi = False
if not have_gloo and not have_mpi:
raise RuntimeError('One of Gloo or MPI are required for Horovod to run. Check the logs above for more info.')
gpu_operations = os.environ.get('HOROVOD_GPU_OPERATIONS')
if gpu_operations and gpu_operations not in {'NCCL', 'MPI'}:
raise DistutilsError(f'HOROVOD_GPU_OPERATIONS={gpu_operations} is invalid, '
f'supported values are: NCCL, MPI.')
def get_gpu_op_variable(op_variable_name, options):
gpu_operation = os.environ.get(op_variable_name)
if gpu_operations:
if gpu_operation:
raise DistutilsError(f'Cannot specify both HOROVOD_GPU_OPERATIONS and {op_variable_name} options. '
'Try unsetting one of these variables and reinstalling.')
return gpu_operations
if gpu_operation and gpu_operation not in options:
options_str = ', '.join(options)
raise DistutilsError(f'{op_variable_name}={gpu_operation} is invalid, '
f'supported values are: {options_str}.')
return gpu_operation
gpu_allreduce = get_gpu_op_variable('HOROVOD_GPU_ALLREDUCE', ['NCCL', 'MPI', 'DDL'])
gpu_allgather = get_gpu_op_variable('HOROVOD_GPU_ALLGATHER', ['NCCL', 'MPI'])
gpu_broadcast = get_gpu_op_variable('HOROVOD_GPU_BROADCAST', ['NCCL', 'MPI'])
gpu_alltoall = get_gpu_op_variable('HOROVOD_GPU_ALLTOALL', ['NCCL', 'MPI'])
have_cuda = False
have_rocm = False
gpu_include_dirs = gpu_lib_dirs = gpu_macros = []
if gpu_allreduce or gpu_allgather or gpu_broadcast or gpu_alltoall:
gpu_type = os.environ.get('HOROVOD_GPU', 'CUDA')
if gpu_type == 'CUDA':
have_cuda = True
gpu_include_dirs, gpu_lib_dirs = get_cuda_dirs(build_ext, cpp_flags)
elif gpu_type == 'ROCM':
have_rocm = True
gpu_include_dirs, gpu_lib_dirs, gpu_macros = get_rocm_dirs(build_ext, cpp_flags)
else:
raise DistutilsError("Unknown HOROVOD_GPU type '%s'" % gpu_type)
if gpu_allreduce == 'NCCL' or gpu_allgather == 'NCCL' or gpu_broadcast == 'NCCL' or gpu_alltoall == 'NCCL':
have_nccl = True
nccl_include_dirs, nccl_lib_dirs, nccl_libs = get_nccl_vals(
build_ext, gpu_include_dirs, gpu_lib_dirs, gpu_macros, cpp_flags, have_rocm)
else:
have_nccl = False
nccl_include_dirs = nccl_lib_dirs = nccl_libs = []
if gpu_allreduce == 'DDL':
warnings.warn('DDL backend has been deprecated. Please, start using the NCCL backend '
'by building Horovod with "HOROVOD_GPU_OPERATIONS=NCCL". '
'Will be removed in v0.21.0.',
DeprecationWarning)
have_ddl = True
ddl_include_dirs, ddl_lib_dirs = get_ddl_dirs(build_ext,
gpu_include_dirs,
gpu_lib_dirs, cpp_flags)
else:
have_ddl = False
ddl_include_dirs = ddl_lib_dirs = []
if gpu_allreduce == 'NCCL' \
and (gpu_allgather == 'MPI' or gpu_broadcast == 'MPI' or gpu_alltoall == 'MPI') \
and not os.environ.get('HOROVOD_ALLOW_MIXED_GPU_IMPL'):
raise DistutilsError(
'You should not mix NCCL and MPI GPU due to a possible deadlock.\n'
'If you\'re sure you want to mix them, set the '
'HOROVOD_ALLOW_MIXED_GPU_IMPL environment variable to \'1\'.')
MACROS = [('EIGEN_MPL2_ONLY', 1)]
INCLUDES = ['third_party/HTTPRequest/include',
'third_party/boost/assert/include',
'third_party/boost/config/include',
'third_party/boost/core/include',
'third_party/boost/detail/include',
'third_party/boost/iterator/include',
'third_party/boost/lockfree/include',
'third_party/boost/mpl/include',
'third_party/boost/parameter/include',
'third_party/boost/predef/include',
'third_party/boost/preprocessor/include',
'third_party/boost/static_assert/include',
'third_party/boost/type_traits/include',
'third_party/boost/utility/include',
'third_party/eigen',
'third_party/flatbuffers/include',
'third_party/lbfgs/include']
SOURCES = ['horovod/common/common.cc',
'horovod/common/controller.cc',
'horovod/common/fusion_buffer_manager.cc',
'horovod/common/half.cc',
'horovod/common/logging.cc',
'horovod/common/message.cc',
'horovod/common/operations.cc',
'horovod/common/parameter_manager.cc',
'horovod/common/response_cache.cc',
'horovod/common/stall_inspector.cc',
'horovod/common/thread_pool.cc',
'horovod/common/timeline.cc',
'horovod/common/tensor_queue.cc',
'horovod/common/ops/collective_operations.cc',
'horovod/common/ops/operation_manager.cc',
'horovod/common/optim/bayesian_optimization.cc',
'horovod/common/optim/gaussian_process.cc',
'horovod/common/utils/env_parser.cc'
]
COMPILE_FLAGS = cpp_flags + shlex.split(mpi_flags)
NVCC_COMPILE_FLAGS = []
LINK_FLAGS = link_flags + shlex.split(mpi_flags)
LIBRARY_DIRS = []
LIBRARIES = []
cpu_operation = os.environ.get('HOROVOD_CPU_OPERATIONS')
if cpu_operation:
print('INFO: Set default CPU operation to ' + cpu_operation)
if cpu_operation.upper() == 'MPI':
if not have_mpi:
raise RuntimeError('MPI is not installed, try changing HOROVOD_CPU_OPERATIONS.')
MACROS += [('HOROVOD_CPU_OPERATIONS_DEFAULT', "'M'")]
elif cpu_operation.upper() == 'MLSL':
raise RuntimeError('Intel(R) MLSL was removed. Upgrade to oneCCL and set HOROVOD_CPU_OPERATIONS=CCL.')
elif cpu_operation.upper() == 'CCL':
if not have_ccl:
raise RuntimeError('oneCCL is not installed, try changing HOROVOD_CPU_OPERATIONS.')
MACROS += [('HOROVOD_CPU_OPERATIONS_DEFAULT', "'C'")]
elif cpu_operation.upper() == 'GLOO':
if compile_without_gloo:
raise ValueError('Cannot set both HOROVOD_WITHOUT_GLOO and HOROVOD_CPU_OPERATIONS=GLOO.')
if is_mac:
raise RuntimeError('Cannot compile Gloo on MacOS, try changing HOROVOD_CPU_OPERATIONS.')
elif not have_cmake:
raise RuntimeError('Cannot compile Gloo without CMake, try installing CMake first.')
MACROS += [('HOROVOD_CPU_OPERATIONS_DEFAULT', "'G'")]
if have_mpi:
MACROS += [('HAVE_MPI', '1')]
SOURCES += ['horovod/common/mpi/mpi_context.cc',
'horovod/common/mpi/mpi_controller.cc',
'horovod/common/ops/mpi_operations.cc',
'horovod/common/ops/adasum/adasum_mpi.cc',
'horovod/common/ops/adasum_mpi_operations.cc']
COMPILE_FLAGS += shlex.split(mpi_flags)
LINK_FLAGS += shlex.split(mpi_flags)
if have_gloo:
MACROS += [('HAVE_GLOO', '1')]
INCLUDES += ['third_party/gloo']
SOURCES += ['horovod/common/gloo/gloo_context.cc',
'horovod/common/gloo/gloo_controller.cc',
'horovod/common/gloo/http_store.cc',
'horovod/common/gloo/memory_store.cc',
'horovod/common/ops/gloo_operations.cc']
if have_ccl:
MACROS += [('HAVE_CCL', '1')]
INCLUDES += [ccl_root + '/include/']
SOURCES += ['horovod/common/ops/ccl_operations.cc']
LIBRARY_DIRS += [ccl_root + '/lib/']
LINK_FLAGS += ['-lccl']
if have_cuda:
set_cuda_options(build_ext, COMPILE_FLAGS, MACROS, INCLUDES, SOURCES, have_mpi, LIBRARY_DIRS,
LIBRARIES, NVCC_COMPILE_FLAGS)
INCLUDES += ['horovod/common/ops/cuda']
if have_rocm:
MACROS += [('HAVE_ROCM', '1'), ('HAVE_GPU', '1')] + gpu_macros
INCLUDES += gpu_include_dirs
SOURCES += ['horovod/common/ops/hip_operations.cc',
'horovod/common/ops/gpu_operations.cc']
if have_mpi:
SOURCES += ['horovod/common/ops/mpi_gpu_operations.cc']
LIBRARY_DIRS += gpu_lib_dirs
LIBRARIES += ['hip_hcc']
if have_nccl:
MACROS += [('HAVE_NCCL', '1')]
INCLUDES += nccl_include_dirs
SOURCES += ['horovod/common/ops/nccl_operations.cc']
if have_mpi:
SOURCES += ['horovod/common/ops/adasum_gpu_operations.cc']
LIBRARY_DIRS += nccl_lib_dirs
LIBRARIES += nccl_libs
if have_ddl and have_mpi:
MACROS += [('HAVE_DDL', '1')]
INCLUDES += ddl_include_dirs
SOURCES += ['horovod/common/mpi/ddl_mpi_context_manager.cc',
'horovod/common/ops/ddl_operations.cc']
LIBRARY_DIRS += ddl_lib_dirs
LIBRARIES += ['ddl', 'ddl_pack']
if gpu_allreduce:
MACROS += [('HOROVOD_GPU_ALLREDUCE', "'%s'" % gpu_allreduce[0])]
if gpu_allgather:
MACROS += [('HOROVOD_GPU_ALLGATHER', "'%s'" % gpu_allgather[0])]
if gpu_broadcast:
MACROS += [('HOROVOD_GPU_BROADCAST', "'%s'" % gpu_broadcast[0])]
if gpu_alltoall:
MACROS += [('HOROVOD_GPU_ALLTOALL', "'%s'" % gpu_alltoall[0])]
return dict(MACROS=MACROS,
INCLUDES=INCLUDES,
SOURCES=SOURCES,
COMPILE_FLAGS=COMPILE_FLAGS,
NVCC_COMPILE_FLAGS=NVCC_COMPILE_FLAGS,
LINK_FLAGS=LINK_FLAGS,
LIBRARY_DIRS=LIBRARY_DIRS,
LIBRARIES=LIBRARIES,
BUILD_GLOO=have_gloo,
BUILD_MPI=have_mpi,
)
def enumerate_binaries_in_path():
for path_dir in os.getenv('PATH', '').split(':'):
if os.path.isdir(path_dir):
for bin_file in sorted(os.listdir(path_dir)):
yield path_dir, bin_file
def determine_gcc_version(compiler):
try:
compiler_macros = subprocess.check_output(
'%s -dM -E - </dev/null' % compiler,
shell=True, universal_newlines=True).split('\n')
for m in compiler_macros:
version_match = re.match('^#define __VERSION__ "(.*?)"$', m)
if version_match:
return LooseVersion(version_match.group(1))
print('INFO: Unable to determine version of the compiler %s.' % compiler)
except subprocess.CalledProcessError:
print('INFO: Unable to determine version of the compiler %s.\n%s'
'' % (compiler, traceback.format_exc()))
return None
def find_gxx_compiler_in_path():
compilers = []
for path_dir, bin_file in enumerate_binaries_in_path():
if re.match('^g\\+\\+(?:-\\d+(?:\\.\\d+)*)?$', bin_file):
# g++, or g++-7, g++-4.9, or g++-4.8.5
compiler = os.path.join(path_dir, bin_file)
compiler_version = determine_gcc_version(compiler)
if compiler_version:
compilers.append((compiler, compiler_version))
return compilers
def find_matching_gcc_compiler_path(gxx_compiler_version):
for path_dir, bin_file in enumerate_binaries_in_path():
if re.match('^gcc(?:-\\d+(?:\\.\\d+)*)?$', bin_file):
# gcc, or gcc-7, gcc-4.9, or gcc-4.8.5
compiler = os.path.join(path_dir, bin_file)
compiler_version = determine_gcc_version(compiler)
if compiler_version and compiler_version == gxx_compiler_version:
return compiler
print('INFO: Unable to find gcc compiler (version %s).' % gxx_compiler_version)
return None
def remove_offensive_gcc_compiler_options(compiler_version):
offensive_replacements = dict()
if compiler_version < LooseVersion('4.9'):
offensive_replacements = {
'-Wdate-time': '',
'-fstack-protector-strong': '-fstack-protector'
}
if offensive_replacements:
from sysconfig import get_config_var
cflags = get_config_var('CONFIGURE_CFLAGS')
cppflags = get_config_var('CONFIGURE_CPPFLAGS')
ldshared = get_config_var('LDSHARED')
for k, v in offensive_replacements.items():
if cflags:
cflags = cflags.replace(k, v)
if cppflags:
cppflags = cppflags.replace(k, v)
if ldshared:
ldshared = ldshared.replace(k, v)
return cflags, cppflags, ldshared
# Use defaults
return None, None, None
# Filter out all the compiler macros (starts with -D)
# that need to be passed to compiler
def filter_compile_macros(compile_flags):
res = []
for flag in compile_flags:
if flag.startswith('-D'):
res.append(flag)
return res
def build_tf_extension(build_ext, global_options):
# Backup the options, preventing other plugins access libs that
# compiled with compiler of this plugin
options = deepcopy(global_options)
tf_version = check_tf_version()
tf_compile_flags, tf_link_flags = get_tf_flags(
build_ext, options['COMPILE_FLAGS'])
gloo_compile_macros = filter_compile_macros(tf_compile_flags)
# Export TENSORFLOW_VERSION equal to our representation of tf.__version__. Internally it's
# used for backwards compatibility checks.
updated_macros = set_macro(
options['MACROS'], 'TENSORFLOW_VERSION', str(tf_version))
tensorflow_mpi_lib.define_macros = updated_macros
tensorflow_mpi_lib.include_dirs = options['INCLUDES']
tensorflow_mpi_lib.sources = options['SOURCES'] + \
['horovod/tensorflow/mpi_ops.cc']
tensorflow_mpi_lib.extra_compile_args = options['COMPILE_FLAGS'] + \
tf_compile_flags
tensorflow_mpi_lib.extra_link_args = options['LINK_FLAGS'] + tf_link_flags
tensorflow_mpi_lib.library_dirs = options['LIBRARY_DIRS']
tensorflow_mpi_lib.libraries = options['LIBRARIES']
cc_compiler = cxx_compiler = cflags = cppflags = ldshared = None
if sys.platform.startswith('linux') and not os.getenv('CC') and not os.getenv('CXX'):