fft_bench.py 4.14 KB
Newer Older
Martin Reinecke's avatar
Martin Reinecke committed
1
import numpy as np
Martin Reinecke's avatar
Martin Reinecke committed
2
import ducc_0_1.fft as duccfft
Martin Reinecke's avatar
Martin Reinecke committed
3
4
5
from time import time
import matplotlib.pyplot as plt

Martin Reinecke's avatar
Martin Reinecke committed
6
7
np.random.seed(42)

Martin Reinecke's avatar
Martin Reinecke committed
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

def _l2error(a, b):
    return np.sqrt(np.sum(np.abs(a-b)**2)/np.sum(np.abs(a)**2))


def measure_fftw(a, nrepeat, nthr, flags=('FFTW_MEASURE',)):
    import pyfftw
    f1 = pyfftw.empty_aligned(a.shape, dtype=a.dtype)
    f2 = pyfftw.empty_aligned(a.shape, dtype=a.dtype)
    fftw = pyfftw.FFTW(f1, f2, flags=flags, axes=range(a.ndim), threads=nthr)
    f1[()] = a
    tmin = 1e38
    for i in range(nrepeat):
        t0 = time()
        fftw()
        t1 = time()
        tmin = min(tmin, t1-t0)
    return tmin, f2


def measure_fftw_est(a, nrepeat, nthr):
    return measure_fftw(a, nrepeat, nthr, flags=('FFTW_ESTIMATE',))


def measure_fftw_np_interface(a, nrepeat, nthr):
    import pyfftw
    pyfftw.interfaces.cache.enable()
    tmin = 1e38
    for i in range(nrepeat):
        t0 = time()
        b = pyfftw.interfaces.numpy_fft.fftn(a)
        t1 = time()
        tmin = min(tmin, t1-t0)
    return tmin, b


Martin Reinecke's avatar
Martin Reinecke committed
44
def measure_duccfft(a, nrepeat, nthr):
Martin Reinecke's avatar
Martin Reinecke committed
45
46
47
48
    tmin = 1e38
    b = a.copy()
    for i in range(nrepeat):
        t0 = time()
Martin Reinecke's avatar
Martin Reinecke committed
49
        b = duccfft.c2c(a, out=b, forward=True, nthreads=nthr)
Martin Reinecke's avatar
Martin Reinecke committed
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
        t1 = time()
        tmin = min(tmin, t1-t0)
    return tmin, b


def measure_scipy_fftpack(a, nrepeat, nthr):
    import scipy.fftpack
    tmin = 1e38
    if nthr!=1:
        raise NotImplementedError("scipy.fftpack does not support multiple threads")
    for i in range(nrepeat):
        t0 = time()
        b = scipy.fftpack.fftn(a)
        t1 = time()
        tmin = min(tmin, t1-t0)
    return tmin, b


def measure_scipy_fft(a, nrepeat, nthr):
    import scipy.fft
    tmin = 1e38
    for i in range(nrepeat):
        t0 = time()
        b = scipy.fft.fftn(a, workers=nthr)
        t1 = time()
        tmin = min(tmin, t1-t0)
    return tmin, b


def measure_numpy_fft(a, nrepeat, nthr):
    import scipy.fft
    tmin = 1e38
    if nthr!=1:
        raise NotImplementedError("numpy.fft does not support multiple threads")
    for i in range(nrepeat):
        t0 = time()
        b = np.fft.fftn(a)
        t1 = time()
        tmin = min(tmin, t1-t0)
    return tmin, b


def measure_mkl_fft(a, nrepeat, nthr):
    import os
    os.environ['OMP_NUM_THREADS'] = str(nthr)
    import mkl_fft
    tmin = 1e38
    for i in range(nrepeat):
        t0 = time()
        b = mkl_fft.fftn(a)
        t1 = time()
        tmin = min(tmin, t1-t0)
    return tmin, b


def bench_nd(ndim, nmax, nthr, ntry, tp, funcs, nrepeat, ttl="", filename="",
             nice_sizes=True):
    print("{}D, type {}, max extent is {}:".format(ndim, tp, nmax))
    results = [[] for i in range(len(funcs))]
    for n in range(ntry):
        shp = np.random.randint(nmax//3, nmax+1, ndim)
        if nice_sizes:
Martin Reinecke's avatar
Martin Reinecke committed
112
            shp = np.array([duccfft.good_size(sz) for sz in shp])
Martin Reinecke's avatar
Martin Reinecke committed
113
114
115
116
117
118
119
        print("  {0:4d}/{1}: shape={2} ...".format(n, ntry, shp), end=" ", flush=True)
        a = (np.random.rand(*shp)-0.5 + 1j*(np.random.rand(*shp)-0.5)).astype(tp)
        output=[]
        for func, res in zip(funcs, results):
            tmp = func(a, nrepeat, nthr)
            res.append(tmp[0])
            output.append(tmp[1])
120
121
122
123
124
125
126
127
128
129
130
131
        print("{0:5.2e}/{1:5.2e} = {2:5.2f}  L2 error={3}".format(results[0][n],results[1][n],results[0][n]/results[1][n],_l2error(output[0],output[1])))
    results = np.array(results)
    plt.title("{}: {}D, {}, max_extent={}".format(
        ttl, ndim, str(tp), nmax))
    plt.xlabel("time ratio")
    plt.ylabel("counts")
    plt.hist(results[0, :]/results[1, :], bins="auto")
    if filename != "":
        plt.savefig(filename)
    plt.show()


Martin Reinecke's avatar
Martin Reinecke committed
132
133
funcs = (measure_duccfft, measure_fftw)
ttl = "duccfft/FFTW()"
Martin Reinecke's avatar
Martin Reinecke committed
134
ntry=100
Martin Reinecke's avatar
Martin Reinecke committed
135
136
nthr = 1
nice_sizes = True
137
bench_nd(1, 8192, nthr, ntry, "c16", funcs, 10, ttl, "1d.png", nice_sizes)
Martin Reinecke's avatar
Martin Reinecke committed
138
bench_nd(2, 2048, nthr, ntry, "c16", funcs, 2, ttl, "2d.png", nice_sizes)
139
140
141
142
bench_nd(3, 256, nthr, ntry, "c16", funcs, 2, ttl, "3d.png", nice_sizes)
bench_nd(1, 8192, nthr, ntry, "c8", funcs, 10, ttl, "1d_single.png", nice_sizes)
bench_nd(2, 2048, nthr, ntry, "c8", funcs, 2, ttl, "2d_single.png", nice_sizes)
bench_nd(3, 256, nthr, ntry, "c8", funcs, 2, ttl, "3d_single.png", nice_sizes)