1
/* Copyright (c) 2011 Julien Pommier ( pommier@modartt.com )
3
Based on original fortran 77 code from FFTPACKv4 from NETLIB,
4
authored by Dr Paul Swarztrauber of NCAR, in 1985.
6
As confirmed by the NCAR fftpack software curators, the following
7
FFTPACKv5 license applies to FFTPACKv4 sources. My changes are
8
released under the same terms.
12
http://www.cisl.ucar.edu/css/software/fftpack5/ftpk.html
14
Copyright (c) 2004 the University Corporation for Atmospheric
15
Research ("UCAR"). All rights reserved. Developed by NCAR's
16
Computational and Information Systems Laboratory, UCAR,
19
Redistribution and use of the Software in source and binary forms,
20
with or without modification, is permitted provided that the
21
following conditions are met:
23
- Neither the names of NCAR's Computational and Information Systems
24
Laboratory, the University Corporation for Atmospheric Research,
25
nor the names of its sponsors or contributors may be used to
26
endorse or promote products derived from this Software without
27
specific prior written permission.
29
- Redistributions of source code must retain the above copyright
30
notices, this list of conditions, and the disclaimer below.
32
- Redistributions in binary form must reproduce the above copyright
33
notice, this list of conditions, and the disclaimer below in the
34
documentation and/or other materials provided with the
37
THIS SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
38
EXPRESS OR IMPLIED, INCLUDING, BUT NOT LIMITED TO THE WARRANTIES OF
39
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
40
NONINFRINGEMENT. IN NO EVENT SHALL THE CONTRIBUTORS OR COPYRIGHT
41
HOLDERS BE LIABLE FOR ANY CLAIM, INDIRECT, INCIDENTAL, SPECIAL,
42
EXEMPLARY, OR CONSEQUENTIAL DAMAGES OR OTHER LIABILITY, WHETHER IN AN
43
ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
44
CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH THE
49
PFFFT : a Pretty Fast FFT.
51
This is basically an adaptation of the single precision fftpack
52
(v4) as found on netlib taking advantage of SIMD instruction found
53
on cpus such as intel x86 (SSE1), powerpc (Altivec), and arm (NEON).
55
For architectures where no SIMD instruction is available, the code
56
falls back to a scalar version.
60
- 1D transforms only, with 32-bit single precision.
62
- supports only transforms for inputs of length N of the form
63
N=(2^a)*(3^b), a >= 5 and b >=0 (32, 48, 64, 96, 128, 144 etc
64
are all acceptable lengths). Performance is best for 128<=N<=8192.
66
- all (float*) pointers in the functions below are expected to
67
have an "simd-compatible" alignment, that is 16 bytes on x86 and
70
You can allocate such buffers with the functions
71
pffft_aligned_malloc / pffft_aligned_free (or with stuff like
85
/* opaque struct holding internal stuff (precomputed twiddle factors)
86
this struct can be shared by many threads as it contains only
89
typedef struct PFFFT_Setup PFFFT_Setup;
91
/* direction of the transform */
92
typedef enum { PFFFT_FORWARD, PFFFT_BACKWARD } pffft_direction_t;
94
/* type of transform */
95
typedef enum { PFFFT_REAL, PFFFT_COMPLEX } pffft_transform_t;
98
prepare for performing transforms of size N -- the returned
99
PFFFT_Setup structure is read-only so it can safely be shared by
100
multiple concurrent threads.
102
static PFFFT_Setup *pffft_new_setup(int N, pffft_transform_t transform);
103
static void pffft_destroy_setup(PFFFT_Setup *);
105
Perform a Fourier transform , The z-domain data is stored in the
106
most efficient order for transforming it back, or using it for
107
convolution. If you need to have its content sorted in the
108
"usual" way, that is as an array of interleaved complex numbers,
109
either use pffft_transform_ordered , or call pffft_zreorder after
110
the forward fft, and before the backward fft.
112
Transforms are not scaled: PFFFT_BACKWARD(PFFFT_FORWARD(x)) = N*x.
113
Typically you will want to scale the backward transform by 1/N.
115
The 'work' pointer should point to an area of N (2*N for complex
116
fft) floats, properly aligned. [del]If 'work' is NULL, then stack will
117
be used instead (this is probably the beest strategy for small
118
FFTs, say for N < 16384).[/del]
120
input and output may alias.
122
static void pffft_transform(PFFFT_Setup *setup, const float *input, float *output, float *work, pffft_direction_t direction);
125
Similar to pffft_transform, but makes sure that the output is
126
ordered as expected (interleaved complex numbers). This is
127
similar to calling pffft_transform and then pffft_zreorder.
129
input and output may alias.
131
static void pffft_transform_ordered(PFFFT_Setup *setup, const float *input, float *output, float *work, pffft_direction_t direction);
134
call pffft_zreorder(.., PFFFT_FORWARD) after pffft_transform(...,
135
PFFFT_FORWARD) if you want to have the frequency components in
136
the correct "canonical" order, as interleaved complex numbers.
138
(for real transforms, both 0-frequency and half frequency
139
components, which are real, are assembled in the first entry as
140
F(0)+i*F(n/2+1). Note that the original fftpack did place
141
F(n/2+1) at the end of the arrays).
143
input and output should not alias.
145
static void pffft_zreorder(PFFFT_Setup *setup, const float *input, float *output, pffft_direction_t direction);
148
Perform a multiplication of the frequency components of dft_a and
149
dft_b and accumulate them into dft_ab. The arrays should have
150
been obtained with pffft_transform(.., PFFFT_FORWARD) and should
151
*not* have been reordered with pffft_zreorder (otherwise just
152
perform the operation yourself as the dft coefs are stored as
153
interleaved complex numbers).
155
the operation performed is: dft_ab += (dft_a * fdt_b)*scaling
157
The dft_a, dft_b and dft_ab pointers may alias.
158
void pffft_zconvolve_accumulate(PFFFT_Setup *setup, const float *dft_a, const float *dft_b, float *dft_ab, float scaling);
162
the operation performed is: dft_ab = (dft_a * fdt_b)
164
The dft_a, dft_b and dft_ab pointers may alias.
166
static void pffft_zconvolve(PFFFT_Setup *setup, const float *dft_a, const float *dft_b, float *dft_ab);
168
/* return 4 or 1 wether support SSE/Altivec instructions was enable when building pffft.c */
169
int pffft_simd_size(void);
171
static void pffft_reorder_back(int length, void * setup, float * data, float * work);