/
cuda.py
640 lines (485 loc) · 19.7 KB
/
cuda.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
"""Device, context and memory management on CuPy.
Chainer uses `CuPy <https://cupy.chainer.org/>`_ (with very thin wrapper)
to exploit the speed of GPU computation. Following modules and classes defined
in CuPy are imported to :mod:`chainer.cuda` module for convenience (refer to
this table when reading chainer's source codes).
============================ =================================
imported name original name
============================ =================================
``chainer.cuda.cupy`` :mod:`cupy`
``chainer.cuda.cupyx`` :mod:`cupyx`
``chainer.cuda.ndarray`` :class:`cupy.ndarray`
``chainer.cuda.cupy.cuda`` :mod:`cupy.cuda`
``chainer.cuda.Device`` :class:`cupy.cuda.Device`
``chainer.cuda.Event`` :class:`cupy.cuda.Event`
``chainer.cuda.Stream`` :class:`cupy.cuda.Stream`
============================ =================================
Chainer replaces the default allocator of CuPy by its memory pool
implementation. It enables us to reuse the device memory over multiple
forward/backward computations, and temporary arrays for consecutive elementwise
operations.
"""
import functools
import os
import warnings
import numpy
import six
import chainer
from chainer.backends import intel64
from chainer.configuration import config
available = False
cudnn_enabled = False
try:
import cupy
from cupy import cuda # NOQA
from cupy.cuda import cublas # NOQA
import cupyx # NOQA
from cupy import ndarray # NOQA
from cupy.cuda import Device # NOQA
from cupy.cuda import Event # NOQA
from cupy.cuda import Stream # NOQA
from . import cuda_fusion as fusion # NOQA
available = True
except Exception as e:
_resolution_error = e
fusion = numpy
class ndarray(object):
pass # for type testing
# for `xp is cuda.cupy` to always work
cupy = object()
if available:
_cudnn_disabled_by_user = int(os.environ.get('CHAINER_CUDNN', '1')) == 0
try:
import cupy.cudnn
cudnn = cupy.cudnn
cudnn_enabled = not _cudnn_disabled_by_user
except Exception as e:
_resolution_error = e
def check_cuda_available():
"""Checks if CUDA is available.
When CUDA is correctly set up, nothing happens.
Otherwise it raises ``RuntimeError``.
"""
if not available:
msg = ('CUDA environment is not correctly set up\n'
'(see https://github.com/chainer/chainer#installation).')
msg += str(_resolution_error)
raise RuntimeError(msg)
if (not cudnn_enabled and
not _cudnn_disabled_by_user and
not getattr(check_cuda_available, '_already_warned', False)):
warnings.warn(
'cuDNN is not enabled.\n'
'Please reinstall CuPy after you install cudnn\n'
'(see https://docs-cupy.chainer.org/en/stable/install.html'
'#install-cupy-with-cudnn-and-nccl).')
check_cuda_available._already_warned = True
class DummyDeviceType(object):
"""Dummy device class that does nothing with cupy.cuda.Device interface.
This class is used to represent CPU device.
"""
id = -1
def __int__(self):
return -1
def __enter__(self):
return self
def __exit__(self, *args):
pass
def use(self):
pass
def synchronize(self):
pass
def __eq__(self, other):
return isinstance(other, DummyDeviceType)
def __ne__(self, other):
return not (self == other)
DummyDevice = DummyDeviceType()
# ------------------------------------------------------------------------------
# Global states
# ------------------------------------------------------------------------------
if available:
# This is for backward compatibility
memory_pool = cupy.get_default_memory_pool()
pinned_memory_pool = cupy.get_default_pinned_memory_pool()
if six.PY2:
try:
from future.types.newint import newint as _newint
_integer_types = six.integer_types + (_newint,)
except ImportError:
_integer_types = six.integer_types
else:
_integer_types = six.integer_types
# ------------------------------------------------------------------------------
# Global states
# ------------------------------------------------------------------------------
def get_device_from_id(device_id):
"""Gets the device from an ID integer.
Args:
device_id (int or None): The ID of the device which this function
returns.
"""
if device_id is not None:
check_cuda_available()
return Device(device_id)
else:
return DummyDevice
def get_device_from_array(*arrays):
"""Gets the device from a list of CuPy array or a single CuPy array.
The device on which the given CuPy array reside is returned.
Args:
array (cupy.ndarray or list of cupy.ndarray):
A CuPy array which this function returns the device corresponding
to. If a list of :class:`cupy.ndarray`\\ s are given, it returns
the first device object of an array in the list.
"""
for array in arrays:
if isinstance(array, ndarray) and array.device is not None:
return array.device
return DummyDevice
def get_device(*args):
"""Gets the device from a device object, an ID integer or an array object.
.. note::
This API is deprecated. Please use
:func:`~chainer.cuda.get_device_from_id`
or :func:`~chainer.cuda.get_device_from_array` instead.
This is a convenient utility to select a correct device if the type of
``arg`` is unknown (i.e., one can use this function on arrays that may be
on CPU or GPU). The returned device object supports the context management
protocol of Python for the *with* statement.
Args:
args: Values to specify a GPU device. The first device object, integer
or :class:`cupy.ndarray` object is used to select a device.
If it is a device object, it is returned. If it is an integer,
the corresponding device is returned. If it is a CuPy array,
the device on which this array reside is returned. If any
arguments are neither integers nor CuPy arrays, a dummy device
object representing CPU is returned.
Returns:
Device object specified by given ``args``.
.. seealso::
See :class:`cupy.cuda.Device` for the device selection not by arrays.
"""
warnings.warn('get_device is deprecated. Please use get_device_from_id or'
' get_device_from_array instead.', DeprecationWarning)
return _get_device(*args)
def _get_device(*args):
for arg in args:
if type(arg) in _integer_types:
check_cuda_available()
return Device(arg)
if isinstance(arg, ndarray):
if arg.device is None:
continue
return arg.device
if available and isinstance(arg, Device):
return arg
return DummyDevice
# ------------------------------------------------------------------------------
# cupy.ndarray allocation and copy
# ------------------------------------------------------------------------------
def to_gpu(array, device=None, stream=None):
"""Copies the given CPU array to the specified device.
Args:
array (*array*, None, list or tuple):
Array or arrays to be sent to GPU.
device: Device specifier.
stream (~cupy.cuda.Stream): *(deprecated since v3.0.0)*
CUDA stream. If not ``None``, the copy runs asynchronously.
Returns:
cupy.ndarray, list or tuple: Array or arrays on GPU.
If some of the arrays are already on GPU, then this function just
returns those arrays without performing any copy.
If input arrays include `None`, it is returned as `None` as is.
"""
if stream is not None:
warnings.warn(
'The stream option is deprecated in chainer.cuda.to_gpu. '
'Please remove it.', DeprecationWarning)
check_cuda_available()
with _get_device(device) as device_:
if isinstance(array, (list, tuple)):
d = {}
ret = []
for arr in array:
if arr is None:
ret.append(None)
else:
arr2 = d.get(id(arr))
if arr2 is None:
arr2 = _array_to_gpu(arr, device_, stream)
d[id(arr)] = arr2
ret.append(arr2)
return type(array)(ret)
else:
return _array_to_gpu(array, device_, stream)
def _array_to_gpu(array, device, stream):
assert device is DummyDevice or isinstance(device, Device)
if array is None:
return None
if isinstance(array, (numpy.number, numpy.bool_)):
array = numpy.asarray(array)
elif isinstance(array, intel64.mdarray):
array = numpy.asarray(array)
if not isinstance(array, (cupy.ndarray, numpy.ndarray)):
raise TypeError(
'The array sent to gpu must be an array or a NumPy scalar.'
'\nActual type: {0}.'.format(type(array)))
array_dev = get_device_from_array(array)
if array_dev.id == cupy.cuda.device.get_device_id():
return array
if stream is not None and stream.ptr != 0:
ret = cupy.empty_like(array)
if array_dev.id == -1:
# cpu to gpu
mem = cupy.cuda.alloc_pinned_memory(array.nbytes)
src = numpy.frombuffer(
mem, array.dtype, array.size).reshape(array.shape)
src[...] = array
ret.set(src, stream)
cupy.cuda.pinned_memory._add_to_watch_list(
stream.record(), mem)
else:
# gpu to gpu
with array_dev:
src = array.copy()
event = Stream.null.record()
stream.wait_event(event)
ret.data.copy_from_device_async(
src.data, src.nbytes, stream)
# to hold a reference until the end of the asynchronous
# memcpy
stream.add_callback(lambda *x: None, (src, ret))
return ret
if array_dev.id == -1:
return cupy.asarray(array)
# Need to make a copy when an array is copied to another device
return cupy.array(array, copy=True)
def to_cpu(array, stream=None):
"""Copies the given GPU array to host CPU.
Args:
array (*array*, None, list or tuple):
Array or arrays to be sent to CPU.
stream (cupy.cuda.Stream): CUDA stream.
Returns:
numpy.ndarray, list or tuple: Array on CPU.
If some of the arrays are already on CPU, then this function just
returns those arrays without performing any copy.
If input arrays include `None`, it is returned as `None` as is.
"""
if isinstance(array, (list, tuple)):
d = {}
ret = []
for arr in array:
if arr is None:
ret.append(None)
else:
arr2 = d.get(id(arr))
if arr2 is None:
arr2 = _array_to_cpu(arr, stream)
d[id(arr)] = arr2
ret.append(arr2)
return type(array)(ret)
else:
return _array_to_cpu(array, stream)
def _array_to_cpu(array, stream):
if array is None:
return None
if isinstance(array, ndarray):
check_cuda_available()
with get_device_from_array(array):
return array.get(stream)
elif isinstance(array, (numpy.number, numpy.bool_)):
return numpy.asarray(array)
elif isinstance(array, chainer.get_cpu_array_types()):
return array
else:
raise TypeError(
'The array sent to cpu must be numpy.ndarray or cupy.ndarray, '
'or a NumPy scalar.'
'\nActual type: {0}.'.format(type(array)))
def copy(array, out=None, out_device=None, stream=None):
"""Copies a :class:`cupy.ndarray` object using the default stream.
This function can copy the device array to the destination array on another
device.
Args:
array (cupy.ndarray): Array to be copied.
out (cupy.ndarray): Destination array.
If it is not ``None``, then ``out_device`` argument is ignored.
out_device: Destination device specifier. Actual device object is
obtained by passing this value to :func:`get_device`.
stream (cupy.cuda.Stream): CUDA stream.
Returns:
cupy.ndarray: Copied array.
If ``out`` is not specified, then the array is allocated on the device
specified by ``out_device`` argument.
"""
check_cuda_available()
assert stream is None # TODO(beam2d): FIX IT
if out is None:
if out_device is None:
out_device = array
with _get_device(out_device):
out = cupy.empty_like(array)
with get_device_from_array(array):
cupy.copyto(out, array)
return out
# ------------------------------------------------------------------------------
# Function result memoization
# ------------------------------------------------------------------------------
def memoize(for_each_device=False):
"""Makes a function memoizing the result for each argument and device.
This is a similar version of :func:`cupy.memoize`. The difference is that
this function can be used in the global scope even if CUDA is not
available. In such case, this function does nothing.
.. note::
This decorator acts as a dummy if CUDA is not available. It cannot be
used for general purpose memoization even if ``for_each_device`` is set
to False.
"""
if available:
return cupy.memoize(for_each_device)
def dummy_decorator(f):
@functools.wraps(f)
def ret(*args, **kwargs):
return f(*args, **kwargs)
return ret
return dummy_decorator
def clear_memo():
"""Clears the memoized results for all functions decorated by memoize.
This function works like :func:`cupy.clear_memo` as a counterpart for
:func:`chainer.cuda.memoize`. It can be used even if CUDA is not available.
In such a case, this function does nothing.
"""
if available:
cupy.clear_memo()
# ------------------------------------------------------------------------------
# Kernel definition utility
# ------------------------------------------------------------------------------
@memoize(for_each_device=True)
def elementwise(in_params, out_params, operation, name, **kwargs):
"""Creates an elementwise kernel function.
This function uses :func:`~chainer.cuda.memoize` to cache the
kernel object, i.e. the resulting kernel object is cached for each argument
combination and CUDA device.
The arguments are the same as those for
:class:`cupy.ElementwiseKernel`, except that the ``name`` argument is
mandatory.
"""
check_cuda_available()
return cupy.ElementwiseKernel(
in_params, out_params, operation, name, **kwargs)
@memoize(for_each_device=True)
def reduce(in_params, out_params, map_expr, reduce_expr, post_map_expr,
identity, name, **kwargs):
"""Creates a global reduction kernel function.
This function uses :func:`~chainer.cuda.memoize` to cache the resulting
kernel object, i.e. the resulting kernel object is cached for each argument
combination and CUDA device.
The arguments are the same as those for
:class:`cupy.ReductionKernel`, except that the ``name`` argument is
mandatory.
"""
check_cuda_available()
return cupy.ReductionKernel(
in_params, out_params, map_expr, reduce_expr, post_map_expr,
identity, name, **kwargs)
# ------------------------------------------------------------------------------
# numpy/cupy compatible coding
# ------------------------------------------------------------------------------
def get_array_module(*args):
"""Gets an appropriate one from :mod:`numpy` or :mod:`cupy`.
This is almost equivalent to :func:`cupy.get_array_module`. The differences
are that this function can be used even if CUDA is not available and that
it will return their data arrays' array module for
:class:`~chainer.Variable` arguments.
Args:
args: Values to determine whether NumPy or CuPy should be used.
Returns:
module: :mod:`cupy` or :mod:`numpy` is returned based on the types of
the arguments.
"""
if available:
args = [arg.data if isinstance(arg, chainer.variable.Variable) else arg
for arg in args]
return cupy.get_array_module(*args)
else:
return numpy
_max_workspace_size = 8 * 1024 * 1024
def get_max_workspace_size():
"""Gets the workspace size for cuDNN.
Check "cuDNN Library User Guide" for detail.
Returns:
int: The workspace size for cuDNN.
"""
return _max_workspace_size
def set_max_workspace_size(size):
"""Sets the workspace size for cuDNN.
Check "cuDNN Library User Guide" for detail.
Args:
size: The workspace size for cuDNN.
"""
global _max_workspace_size
_max_workspace_size = size
def fuse(*args, **kwargs):
"""Function fusing decorator.
It calls :func:`cupy.fuse` when CuPy is available to make fused function
and does nothing otherwise.
.. seealso::
:func:`cupy.fuse`
"""
if available:
return cupy.fuse(*args, **kwargs)
else:
return lambda f: f
# ------------------------------------------------------------------------------
# cuDNN
# ------------------------------------------------------------------------------
_SHOULD_USE_CUDNN = {
'==always': {'always': True, 'auto': False, 'never': False},
'>=auto': {'always': True, 'auto': True, 'never': False},
}
_cudnn_version = cuda.cudnn.getVersion() if cudnn_enabled else -1
def should_use_cudnn(level, lowest_version=0):
"""Determines if we should use cuDNN.
This function checks ``chainer.config.use_cudnn``,
``chainer.cuda.cudnn_enabled``, and the cuDNN version. Note that
``cudnn_enabled`` flag is fixed at loading of :mod:`chainer` module.
Args:
level (str): cuDNN use level. It must be either ``'==always'`` or
``'>=auto'``. ``'==always'`` indicates that the ``use_cudnn``
config must be ``'always'`` to use cuDNN.
lowest_version (int): Required lowest cuDNN version. It must be
non-negative.
Returns:
bool: ``True`` if the caller should use cuDNN.
"""
if _cudnn_version < lowest_version:
return False
if level not in _SHOULD_USE_CUDNN:
raise ValueError('invalid cuDNN use level: %s '
'(must be either of "==always" or ">=auto")' %
repr(level))
flags = _SHOULD_USE_CUDNN[level]
use_cudnn = config.use_cudnn
if use_cudnn not in flags:
raise ValueError('invalid use_cudnn configuration: %s '
'(must be either of "always", "auto", or "never")' %
repr(use_cudnn))
return flags[use_cudnn]
_tensor_core_flag = {'always': True, 'auto': None, 'never': False}
def should_use_cudnn_tensor_core(dtype):
"""Determines if Tensor Core should be used.
Args:
dtype (numpy.dtype): data type of input tensor.
Returns:
bool: ``True`` if Tensor Core should be used.
"""
use_cudnn_tensor_core = config.use_cudnn_tensor_core
if use_cudnn_tensor_core not in _tensor_core_flag:
raise ValueError('invalid use_cudnn_tensor_core configuration: %s '
'(must be either of "always", "auto", or "never")' %
repr(use_cudnn_tensor_core))
use_tensor_core = _tensor_core_flag[use_cudnn_tensor_core]
if use_tensor_core is None:
use_tensor_core = cudnn.is_tensor_core_available(dtype)
return use_tensor_core