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
|
from time import time
import numpy as np
import pyfftw
import scipy.fftpack as sp
from mpi4py_fft import fftw
import pickle
try:
#fftw.import_wisdom('wisdom.dat')
pyfftw.import_wisdom(pickle.load(open('pyfftw.wisdom', 'rb')))
print('Wisdom imported')
except:
print('Wisdom not imported')
N = (64, 64, 64)
loops = 50
axis = 1
threads = 4
implicit = True
flags = (fftw.FFTW_PATIENT, fftw.FFTW_DESTROY_INPUT)
# Transform complex to complex
#A = pyfftw.byte_align(np.random.random(N).astype('D'))
#A = np.random.random(N).astype(np.dtype('D'))
A = fftw.aligned(N, n=8, dtype=np.dtype('D'))
A[:] = np.random.random(N).astype(np.dtype('D'))
#print(A.ctypes.data % 32)
input_array = fftw.aligned(A.shape, n=32, dtype=A.dtype)
output_array = fftw.aligned(A.shape, n=32, dtype=A.dtype)
AC = A.copy()
ptime = [[], []]
ftime = [[], []]
stime = [[], []]
for axis in ((1, 2), 0, 1, 2):
axes = axis if np.ndim(axis) else [axis]
# pyfftw
fft = pyfftw.builders.fftn(input_array, axes=axes, threads=threads,
overwrite_input=True)
t0 = time()
for i in range(loops):
C = fft(A)
ptime[0].append(time()-t0)
# us
fft = fftw.fftn(input_array, None, axes, threads, flags)
t0 = time()
for i in range(loops):
C2 = fft(A, implicit=implicit)
ftime[0].append(time()-t0)
assert np.allclose(C, C2)
# scipy
if not A.dtype.char.upper() == 'G':
C3 = sp.fftn(A, axes=axes) # scipy is caching, so call once before
t0 = time()
for i in range(loops):
C3 = sp.fftn(A, axes=axes)
stime[0].append(time()-t0)
else:
stime[0].append(0)
# pyfftw
ifft = pyfftw.builders.ifftn(output_array, axes=axes, threads=threads,
overwrite_input=True)
CC = C.copy()
t0 = time()
for i in range(loops):
B = ifft(C, normalise_idft=True)
ptime[1].append(time()-t0)
# us
ifft = fftw.ifftn(output_array, None, axes, threads, flags)
t0 = time()
for i in range(loops):
B2 = ifft(C, normalize=True, implicit=implicit)
ftime[1].append(time()-t0)
assert np.allclose(B, B2), np.linalg.norm(B-B2)
# scipy
if not C.dtype.char.upper() == 'G':
B3 = sp.ifftn(C, axes=axes) # scipy is caching, so call once before
t0 = time()
for i in range(loops):
B3 = sp.ifftn(C, axes=axes)
stime[1].append(time()-t0)
else:
stime[1].append(0)
print("Timing forward transform axes (1, 2), 0, 1, 2")
print("pyfftw {0:2.4e} {1:2.4e} {2:2.4e} {3:2.4e}".format(*ptime[0]))
print("mpi4py {0:2.4e} {1:2.4e} {2:2.4e} {3:2.4e}".format(*ftime[0]))
print("scipy {0:2.4e} {1:2.4e} {2:2.4e} {3:2.4e}".format(*stime[0]))
print("Timing backward transform axes (1, 2), 0, 1, 2")
print("pyfftw {0:2.4e} {1:2.4e} {2:2.4e} {3:2.4e}".format(*ptime[1]))
print("mpi4py {0:2.4e} {1:2.4e} {2:2.4e} {3:2.4e}".format(*ftime[1]))
print("scipy {0:2.4e} {1:2.4e} {2:2.4e} {3:2.4e}".format(*stime[1]))
# Transform real to complex
# Not scipy because they do not have rfftn
#A = pyfftw.byte_align(np.random.random(N).astype('d'))
A = np.random.random(N).astype(np.dtype('d', align=True))
input_array = np.zeros_like(A)
ptime = [[], []]
ftime = [[], []]
for axis in ((1, 2), 0, 1, 2):
axes = axis if np.ndim(axis) else [axis]
# pyfftw
rfft = pyfftw.builders.rfftn(input_array, axes=axes, threads=threads)
t0 = time()
for i in range(loops):
C = rfft(A)
ptime[0].append(time()-t0)
# us
rfft = fftw.rfftn(input_array, None, axes, threads, flags)
t0 = time()
for i in range(loops):
C2 = rfft(A, implicit=implicit)
ftime[0].append(time()-t0)
assert np.allclose(C, C2)
# pyfftw
irfft = pyfftw.builders.irfftn(C.copy(), s=np.take(input_array.shape, axes),
axes=axes, threads=threads)
t0 = time()
for i in range(loops):
C2[:] = C # Because irfft is overwriting input
D = irfft(C2, normalise_idft=True)
ptime[1].append(time()-t0)
# us
irfft = fftw.irfftn(C.copy(), np.take(input_array.shape, axes), axes, threads, flags)
t0 = time()
for i in range(loops):
C2[:] = C
D2 = irfft(C2, normalize=True, implicit=implicit)
ftime[1].append(time()-t0)
assert np.allclose(D, D2), np.linalg.norm(D-D2)
print("Timing real forward transform axes (1, 2), 0, 1, 2")
print("pyfftw {0:2.4e} {1:2.4e} {2:2.4e} {3:2.4e}".format(*ptime[0]))
print("mpi4py {0:2.4e} {1:2.4e} {2:2.4e} {3:2.4e}".format(*ftime[0]))
print("Timing real backward transform axes (1, 2), 0, 1, 2")
print("pyfftw {0:2.4e} {1:2.4e} {2:2.4e} {3:2.4e}".format(*ptime[1]))
print("mpi4py {0:2.4e} {1:2.4e} {2:2.4e} {3:2.4e}".format(*ftime[1]))
fftw.export_wisdom('wisdom.dat')
|
