-
Notifications
You must be signed in to change notification settings - Fork 10
/
fftconv.py
682 lines (525 loc) · 22.7 KB
/
fftconv.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
import string
import numpy as np
import theano
import theano.tensor as T
from theano.sandbox.cuda import cuda_available, GpuOp
from theano.ifelse import ifelse
if cuda_available:
from theano.sandbox.cuda import (basic_ops, CudaNdarrayType,
CudaNdarray)
from theano.misc.pycuda_init import pycuda_available
if pycuda_available:
import pycuda.gpuarray
try:
import scikits.cuda
from scikits.cuda import fft, cublas
scikits.cuda.misc.init()
scikits_cuda_available = True
except ImportError:
scikits_cuda_available = False
# TODO: investigate the effect of enabling fastmath on FFT performance
# (how can it be enabled?).
# base class for shared code between scikits.cuda-based ops
class ScikitsCudaOp(GpuOp):
def __eq__(self, other):
return type(self) == type(other)
def __hash__(self):
return hash(type(self))
def __str__(self):
return self.__class__.__name__
def output_type(self, inp):
raise NotImplementedError
def make_node(self, inp):
inp = basic_ops.gpu_contiguous(
basic_ops.as_cuda_ndarray_variable(inp))
assert inp.dtype == "float32"
return theano.Apply(self, [inp], [self.output_type(inp)()])
def make_thunk(self, node, storage_map, _, _2):
if not scikits_cuda_available:
raise RuntimeError(
"scikits.cuda is needed for all GPU fft implementation,"
" including fftconv.")
class CuFFTOp(ScikitsCudaOp):
def output_type(self, inp):
# add one extra dim for real/imag
return CudaNdarrayType(
broadcastable=[False] * (inp.type.ndim))
def make_thunk(self, node, storage_map, compute_map, no_recycling):
super(CuFFTOp, self).make_thunk(node, storage_map, compute_map, no_recycling)
from theano.misc.pycuda_utils import to_gpuarray
inputs = [storage_map[v] for v in node.inputs]
outputs = [storage_map[v] for v in node.outputs]
plan_input_shape = [None]
plan = [None]
def thunk():
input_shape = inputs[0][0].shape
output_shape = input_shape
z = outputs[0]
# only allocate if there is no previous allocation of the
# right size.
if z[0] is None or z[0].shape != output_shape:
z[0] = CudaNdarray.zeros(output_shape)
input_pycuda = to_gpuarray(inputs[0][0])
# I thought we'd need to change the type on output_pycuda
# so it is complex64, but as it turns out scikits.cuda.fft
# doesn't really care either way and treats the array as
# if it is complex64 anyway.
output_pycuda = to_gpuarray(z[0])
# only initialise plan if necessary
if plan[0] is None or plan_input_shape[0] != input_shape:
plan_input_shape[0] = input_shape
plan[0] = fft.Plan(input_shape[1:-1], np.complex64, np.complex64,
batch=input_shape[0])
fft.fft(input_pycuda, output_pycuda, plan[0])
compute_map[node.outputs[0]][0] = True
thunk.inputs = inputs
thunk.outputs = outputs
thunk.lazy = False
return thunk
def grad(self, inputs, output_grads):
return [CuIFFTOp()(output_grads[0])]
class CuIFFTOp(ScikitsCudaOp):
def output_type(self, inp):
# remove extra real/imag dim
return CudaNdarrayType(
broadcastable=[False] * (inp.type.ndim))
def make_thunk(self, node, storage_map, compute_map, no_recycling):
super(CuIFFTOp, self).make_thunk(node, storage_map, compute_map, no_recycling)
from theano.misc.pycuda_utils import to_gpuarray
inputs = [storage_map[v] for v in node.inputs]
outputs = [storage_map[v] for v in node.outputs]
plan_input_shape = [None]
plan = [None]
def thunk():
input_shape = inputs[0][0].shape
output_shape = input_shape
z = outputs[0]
# only allocate if there is no previous allocation of the
# right size.
if z[0] is None or z[0].shape != output_shape:
z[0] = CudaNdarray.zeros(output_shape)
input_pycuda = to_gpuarray(inputs[0][0])
# input_pycuda is a float32 array with an extra dimension,
# but will be interpreted by scikits.cuda as a complex64
# array instead.
output_pycuda = to_gpuarray(z[0])
# only initialise plan if necessary
if plan[0] is None or plan_input_shape[0] != input_shape:
plan_input_shape[0] = input_shape
plan[0] = fft.Plan(output_shape[1:-1], np.complex64, np.complex64,
batch=output_shape[0])
fft.ifft(input_pycuda, output_pycuda, plan[0])
compute_map[node.outputs[0]][0] = True
# strangely enough, enabling rescaling here makes it run
# very, very slowly. so do this rescaling manually
# afterwards!
thunk.inputs = inputs
thunk.outputs = outputs
thunk.lazy = False
return thunk
def grad(self, inputs, output_grads):
return [CuFFTOp()(output_grads[0])]
def to_complex_gpuarray(x, copyif=False):
"""
Adapted version of theano.misc.pycuda_utils.to_gpuarray that takes
an array with an extra trailing dimension of length 2 for
real/imaginary parts, and turns it into a complex64 PyCUDA
GPUArray.
"""
if not isinstance(x, CudaNdarray):
raise ValueError("We can transfer only CudaNdarray "
"to pycuda.gpuarray.GPUArray")
else:
# Check if trailing dimension has length 2
assert x.shape[-1] == 2
# check if dtype is float32
assert x.dtype == 'float32'
# Check if it is c contiguous
size = 1
c_contiguous = True
for i in range(x.ndim - 1, -1, -1):
if x.shape[i] == 1:
continue
if x._strides[i] != size:
c_contiguous = False
break
size *= x.shape[i]
if not c_contiguous:
if copyif:
x = x.copy()
else:
raise ValueError("We were asked to not copy memory, "
"but the memory is not c contiguous.")
# Now x is always c contiguous
px = pycuda.gpuarray.GPUArray(x.shape[:-1], np.complex64, base=x,
gpudata=x.gpudata)
return px
def bptrs(a):
"""
Pointer array when input represents a batch of matrices.
Taken from scikits.cuda tests/test_cublas.py.
"""
return pycuda.gpuarray.arange(a.ptr, a.ptr + a.shape[0] * a.strides[0],
a.strides[0], dtype=cublas.ctypes.c_void_p)
def sc_complex_dot_batched(bx_gpu, by_gpu, bc_gpu, transa='N', transb='N',
handle=None):
"""
Uses cublasCgemmBatched to compute a bunch of complex dot products
in parallel.
"""
if handle is None:
handle = scikits.cuda.misc._global_cublas_handle
assert len(bx_gpu.shape) == 3
assert len(by_gpu.shape) == 3
assert len(bc_gpu.shape) == 3
assert bx_gpu.dtype == np.complex64
assert by_gpu.dtype == np.complex64
assert bc_gpu.dtype == np.complex64
# Get the shapes of the arguments
bx_shape = bx_gpu.shape
by_shape = by_gpu.shape
# Perform matrix multiplication for 2D arrays:
alpha = np.complex64(1.0)
beta = np.complex64(0.0)
transa = string.lower(transa)
transb = string.lower(transb)
if transb in ['t', 'c']:
N, m, k = by_shape
elif transb in ['n']:
N, k, m = by_shape
else:
raise ValueError('invalid value for transb')
if transa in ['t', 'c']:
N2, l, n = bx_shape
elif transa in ['n']:
N2, n, l = bx_shape
else:
raise ValueError('invalid value for transa')
if l != k:
raise ValueError('objects are not aligned')
if N != N2:
raise ValueError('batch sizes are not the same')
if transb == 'n':
lda = max(1, m)
else:
lda = max(1, k)
if transa == 'n':
ldb = max(1, k)
else:
ldb = max(1, n)
ldc = max(1, m)
# construct pointer arrays needed for cublasCgemmBatched
bx_arr = bptrs(bx_gpu)
by_arr = bptrs(by_gpu)
bc_arr = bptrs(bc_gpu)
cublas.cublasCgemmBatched(handle, transb, transa, m, n, k, alpha,
by_arr.gpudata, lda, bx_arr.gpudata, ldb,
beta, bc_arr.gpudata, ldc, N)
class BatchedComplexDotOp(ScikitsCudaOp):
"""
This version uses cublasCgemmBatched under the hood, instead of
doing multiple cublasCgemm calls.
"""
def make_node(self, inp1, inp2):
inp1 = basic_ops.gpu_contiguous(
basic_ops.as_cuda_ndarray_variable(inp1))
inp2 = basic_ops.gpu_contiguous(
basic_ops.as_cuda_ndarray_variable(inp2))
assert inp1.dtype == "float32"
assert inp2.dtype == "float32"
assert inp1.ndim == 4 # (batch, a, b, real/imag)
assert inp2.ndim == 4
return theano.Apply(self, [inp1, inp2], [self.output_type(inp1)()])
def output_type(self, inp):
return CudaNdarrayType(broadcastable=[False] * inp.type.ndim)
def make_thunk(self, node, storage_map, _, _2):
super(BatchedComplexDotOp, self).make_thunk(node, storage_map, _, _2)
inputs = [storage_map[v] for v in node.inputs]
outputs = [storage_map[v] for v in node.outputs]
def thunk():
bx = inputs[0]
by = inputs[1]
input_shape_x = bx[0].shape # (batch, a, b, 2)
input_shape_y = by[0].shape # (batch, b, c, 2)
output_shape = (input_shape_x[0], input_shape_x[1],
input_shape_y[2], 2) # (batch, a, c, 2)
bz = outputs[0]
# only allocate if there is no previous allocation of the
# right size.
if bz[0] is None or bz[0].shape != output_shape:
bz[0] = CudaNdarray.zeros(output_shape)
input_bx_pycuda = to_complex_gpuarray(bx[0])
input_by_pycuda = to_complex_gpuarray(by[0])
output_b_pycuda = to_complex_gpuarray(bz[0])
# fancy native batched version
sc_complex_dot_batched(input_bx_pycuda, input_by_pycuda,
output_b_pycuda)
thunk.inputs = inputs
thunk.outputs = outputs
thunk.lazy = False
return thunk
cufft = CuFFTOp()
cuifft = CuIFFTOp()
batched_complex_dot = BatchedComplexDotOp()
def mult_and_reduce(input_fft_v, filters_fft_v, input_shape=None,
filter_shape=None):
"""
Parameters
----------
input_fft_v
It's (b, ic, i0, i1//2 + 1, 2).
filters_fft_v
It's (oc, ic, i0, i1//2 + 1, 2).
"""
if input_shape is None:
input_shape = input_fft_v.shape # symbolic
if filter_shape is None:
filter_shape = filters_fft_v.shape # symbolic
b, ic, i0, i1_f, _ = input_shape
oc = filter_shape[0]
# reshape to flatten the dimensions that are multiplied elemwise
input_r = input_fft_v.reshape((b, ic, i0 * i1_f, 2))
filters_r = filters_fft_v.reshape((oc, ic, i0 * i1_f, 2))
# shuffle for batched dot product
input_s = input_r.dimshuffle(2, 0, 1, 3) # (i0 * i1_f, b, ic, 2)
filters_s = filters_r.dimshuffle(2, 1, 0, 3) # (i0 * i1_f, ic, oc, 2)
output_s = batched_complex_dot(input_s, filters_s)
# shuffle again
output_r = output_s.dimshuffle(1, 2, 0, 3)
# reshape to unflatten
output = output_r.reshape((b, oc, i0, i1_f, 2))
return output
def conv2d_fft(input, filters, image_shape=None, filter_shape=None,
border_mode='valid', pad_last_dim=False):
"""
Perform a convolution through fft.
Only support input which will be even on the last dimension
(width). All other dimensions can be anything and the filters can
have an even or odd width.
If you must use input which has an odd width, you can either pad
it or use the `pad_last_dim` argument which will do it for you and
take care to strip the padding before returning. Don't use this
argument if you are not sure the input is odd since the padding is
unconditional and will make even input odd, thus leading to
problems.
On valid mode the filters must be smaller than the input.
Parameters
----------
input
(b, ic, i0, i1).
filters
(oc, ic, f0, f1).
border_mode : {'valid', 'full'}
pad_last_dim
Unconditionally pad the last dimension of the input
to to turn it from odd to even. Will strip the
padding before returning the result.
"""
# use symbolic shapes to compute shape info at runtime if not specified
if image_shape is None:
image_shape = input.shape
if filter_shape is None:
filter_shape = filters.shape
# batch size, input channels, input dim 0, input dim 1
b, ic, i0, i1 = image_shape
# output channels, input channels, filter dim 0, filter dim 1
oc, ic_, f0, f1 = filter_shape
# pad filters/image to output shape
if border_mode == 'valid':
o0 = i0
if pad_last_dim:
o1 = i1 + 1
input_padded = T.zeros((b, ic, o0, o1), dtype='float32')
input_padded = T.set_subtensor(input_padded[:, :, :i0, :i1],
input)
else:
o1 = i1
input_padded = input
filters_padded = T.zeros((oc, ic, o0, o1), dtype='float32')
filters_padded = T.set_subtensor(filters_padded[:, :, :f0, :f1],
filters)
elif border_mode == 'full':
# In this particular case, the values of (o0, o1) represent
# the dimensions of the work buffer more than the actual dimensions
# of the desired output.
o0 = i0 + 2 * (f0 - 1)
o1 = i1 + 2 * (f1 - 1)
if pad_last_dim:
o1 = o1 + 1
# We line up the filters and the images in a way
# such that the filters are tightly placed against the
# top-left of the array, and the images intersect with
# them on one pixel. The top-left pixel of the images
# is the bottom-right pixel of the filters when we
# do the layout here.
filters_padded = T.zeros((oc, ic, o0, o1), dtype='float32')
filters_padded = T.set_subtensor(filters_padded[:, :, :f0, :f1],
filters)
input_padded = T.zeros((b, ic, o0, o1), dtype='float32')
input_padded = T.set_subtensor(input_padded[:, :, (f0 - 1):(f0 - 1 + i0), (f1 - 1):(f1 - 1 + i1)],
input)
else:
raise ValueError('invalid mode')
input_padded = T.opt.Assert("in conv2d_fft: width is not even")(
input_padded, T.eq(o1 % 2, 0))
# reshape for FFT
input_flat = input_padded.reshape((b * ic, o0, o1))
filters_flat = filters_padded.reshape((oc * ic, o0, o1))
# perform FFT
input_fft_flat = cufft(input_flat) # (b * ic, o0, o1//2 + 1, 2)
filters_fft_flat = cufft(filters_flat) # (oc * ic, o0, o1//2 + 1, 2)
# unfold ic dimension
input_fft_v_shape = (b, ic, o0, o1 // 2 + 1, 2)
filters_fft_v_shape = (oc, ic, o0, o1 // 2 + 1, 2)
input_fft_v = input_fft_flat.reshape(input_fft_v_shape)
filters_fft_v = filters_fft_flat.reshape(filters_fft_v_shape)
# (b, oc, o0, o1//2 + 1, 2)
output_fft_s = mult_and_reduce(input_fft_v, filters_fft_v,
input_shape=input_fft_v_shape,
filter_shape=filters_fft_v_shape)
# reshape for IFFT
output_fft_flat = output_fft_s.reshape((b * oc, o0, o1 // 2 + 1, 2))
# perform IFFT
output_flat = cuifft(output_fft_flat) # (b * oc, o0, o1)
# reshape
output_circ = output_flat.reshape((b, oc, o0, o1)) # circular!
# Now we extract the region of interest.
# We just cut it out from the output_circ
# array that was used for the computation.
# We do not need to handle pad_last_dim in a
# special way because we specify explicitly here
# how much values are expected.
if border_mode == 'valid':
output = output_circ[:, :, (f0-1):(f0-1 + i0-f0+1), (f1-1):(f1-1 + i1-f1+1)]
elif border_mode == 'full':
output = output_circ[:, :, (f0-1):(f0-1 + i0+f0-1), (f1-1):(f1-1 + i1+f1-1)]
else:
raise ValueError('invalid mode')
# Rescale manually. This is just a factor that comes in during the
# trip through FFT and inverse FFT.
output = (1.0 / T.cast(o0 * o1, 'float32')) * output
# output should now be the result of a batched valid convolution
# of the input with the filters.
return basic_ops.as_cuda_ndarray_variable(output)
def conv3d_fft(input, filters, image_shape=None, filter_shape=None,
border_mode='valid', pad_last_dim=False):
"""
Perform a convolution through fft.
Only supports input whose shape is even on the last dimension.
All other dimensions can be anything and the filters can
have an even or odd last dimension.
The semantics associated with the last three dimensions
are not important as long as they are in the same order between
the inputs and the filters. For example, when the convolution
is done on a sequence of images, they could be either
(duration, height, width) or (height, width, duration).
If you must use input which has an odd width, you can either pad
it or use the `pad_last_dim` argument which will do it for you and
take care to strip the padding before returning. pad_last_dim checks
that the last dimension is odd before the actual paddding
On valid mode the filters must be smaller than the input.
Parameters
----------
input
(b, ic, i0, i1, i2).
filters
(oc, ic, f0, f1, i2).
border_mode : {'valid', 'full'}.
pad_last_dim
Unconditionally pad the last dimension of the input
to to turn it from odd to even. Will strip the
padding before returning the result.
"""
# use symbolic shapes to compute shape info at runtime if not specified
if image_shape is None:
image_shape = input.shape
if filter_shape is None:
filter_shape = filters.shape
# batch size, input channels, input dim 0, input dim 1
b, ic, i0, i1, i2 = image_shape
# output channels, input channels, filter dim 0, filter dim 1
oc, ic_, f0, f1, f2 = filter_shape
# Check that the last dimension is odd
is_odd = T.eq(T.mod(input.shape[4], 2), 1)
# pad filters/image to output shape
if border_mode == 'valid':
o0 = i0
o1 = i1
o2 = i2
input_padded = input
if pad_last_dim:
o2 = ifelse(is_odd, o2 + 1, o2)
input_padded = T.zeros((b, ic, o0, o1, o2), dtype='float32')
input_padded = T.set_subtensor(input_padded[:, :, :i0, :i1, :i2],
input)
filters_padded = T.zeros((oc, ic, o0, o1, o2), dtype='float32')
filters_padded = T.set_subtensor(filters_padded[:, :, :f0, :f1, :f2],
filters)
elif border_mode == 'full':
# In this particular case, the values of (o0, o1) represent
# the dimensions of the work buffer more than the actual dimensions
# of the desired output.
o0 = i0 + 2 * (f0 - 1)
o1 = i1 + 2 * (f1 - 1)
o2 = i2 + 2 * (f2 - 1)
if pad_last_dim:
o2 = ifelse(is_odd, o2 + 1, o2)
# We line up the filters and the images in a way
# such that the filters are tightly placed against the
# top-left of the array, and the images intersect with
# them on one pixel. The top-left pixel of the images
# is the bottom-right pixel of the filters when we
# do the layout here.
filters_padded = T.zeros((oc, ic, o0, o1, o2), dtype='float32')
filters_padded = T.set_subtensor(filters_padded[:, :, :f0, :f1, :f2],
filters)
input_padded = T.zeros((b, ic, o0, o1, o2), dtype='float32')
input_padded = T.set_subtensor(input_padded[:, :, (f0 - 1):(f0 - 1 + i0), (f1 - 1):(f1 - 1 + i1), (f2 - 1):(f2 - 1 + i2)],
input)
else:
raise ValueError('invalid mode')
# reshape for FFT
input_flat = input_padded.reshape((b * ic, o0, o1, o2))
filters_flat = filters_padded.reshape((oc * ic, o0, o1, o2))
# perform FFT
input_fft_flat = cufft(input_flat) # (b * ic, o0, o1, o2//2 + 1, 2)
filters_fft_flat = cufft(filters_flat) # (oc * ic, o0, o1, o2//2 + 1, 2)
# Unfold ic dimension.
# We have to collapse two dimensions together
# in order to reuse the same `mult_and_reduce`.
# This explains the o0 * 01 instead of just keeping
# the two dimensions intact.
input_fft_v_shape = (b, ic, o0 * o1, o2 // 2 + 1, 2)
filters_fft_v_shape = (oc, ic, o0 * o1, o2 // 2 + 1, 2)
input_fft_v = input_fft_flat.reshape(input_fft_v_shape)
filters_fft_v = filters_fft_flat.reshape(filters_fft_v_shape)
# (b, oc, o0 * o1, o2//2 + 1, 2)
output_fft_s = mult_and_reduce(input_fft_v, filters_fft_v,
input_shape=input_fft_v_shape,
filter_shape=filters_fft_v_shape)
#output_fft_s = input_fft_v
# reshape for IFFT
output_fft_flat = output_fft_s.reshape((b * oc, o0, o1, o2 // 2 + 1, 2))
# perform IFFT
output_flat = cuifft(output_fft_flat) # (b * oc, o0, o1, o2)
# reshape
output_circ = output_flat.reshape((b, oc, o0, o1, o2)) # circular!
# Now we extract the region of interest.
# We just cut it out from the output_circ
# array that was used for the computation.
# We do not need to handle pad_last_dim in a
# special way because we specify explicitly here
# how much values are expected.
if border_mode == 'valid':
output = output_circ[:, :, (f0-1):(f0-1 + i0-f0+1), (f1-1):(f1-1 + i1-f1+1), (f2-1):(f2-1 + i2-f2+1)]
elif border_mode == 'full':
output = output_circ[:, :, (f0-1):(f0-1 + i0+f0-1), (f1-1):(f1-1 + i1+f1-1), (f2-1):(f2-1 + i2+f2-1)]
else:
raise ValueError('invalid mode')
#output = output_circ[:, :, :, :, :]
# Rescale manually. This is just a factor that comes in during the
# trip through FFT and inverse FFT.
output = (1.0 / T.cast(o0 * o1 * o2, 'float32')) * output
# output should now be the result of a batched valid convolution
# of the input with the filters.
return basic_ops.as_cuda_ndarray_variable(output)