remove unused exptab field of struct fft_complex.
[paraslash.git] / imdct.c
1 /*
2 * FFT/IFFT transforms.
3 *
4 * Extracted 2009 from mplayer 2009-02-10 libavcodec/fft.c and libavcodec/mdct.c
5 *
6 * Copyright (c) 2008 Loren Merritt
7 * Copyright (c) 2002 Fabrice Bellard
8 * Partly based on libdjbfft by D. J. Bernstein
9 *
10 * Licensed under the GNU Lesser General Public License.
11 * For licencing details see COPYING.LIB.
12 */
13
14 /**
15 * \file imdct.c Inverse modified discrete cosine transform.
16 */
17
18 #include <inttypes.h>
19 #include <math.h>
20 #include <string.h>
21 #include <stdlib.h>
22 #include <regex.h>
23
24 #include "para.h"
25 #include "error.h"
26 #include "string.h"
27 #include "imdct.h"
28 #include "wma.h"
29
30 typedef float fftsample_t;
31
32 struct fft_complex {
33 fftsample_t re, im;
34 };
35
36 struct fft_context {
37 int nbits;
38 uint16_t *revtab;
39 };
40
41 struct mdct_context {
42 /** Size of MDCT (i.e. number of input data * 2). */
43 int n;
44 /** n = 2^n bits. */
45 int nbits;
46 /** pre/post rotation tables */
47 fftsample_t *tcos;
48 fftsample_t *tsin;
49 struct fft_context fft;
50 };
51
52 /* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */
53 DECLARE_ALIGNED_16(fftsample_t, ff_cos_16[8]);
54 DECLARE_ALIGNED_16(fftsample_t, ff_cos_32[16]);
55 DECLARE_ALIGNED_16(fftsample_t, ff_cos_64[32]);
56 DECLARE_ALIGNED_16(fftsample_t, ff_cos_128[64]);
57 DECLARE_ALIGNED_16(fftsample_t, ff_cos_256[128]);
58 DECLARE_ALIGNED_16(fftsample_t, ff_cos_512[256]);
59 DECLARE_ALIGNED_16(fftsample_t, ff_cos_1024[512]);
60 DECLARE_ALIGNED_16(fftsample_t, ff_cos_2048[1024]);
61 DECLARE_ALIGNED_16(fftsample_t, ff_cos_4096[2048]);
62 DECLARE_ALIGNED_16(fftsample_t, ff_cos_8192[4096]);
63 DECLARE_ALIGNED_16(fftsample_t, ff_cos_16384[8192]);
64 DECLARE_ALIGNED_16(fftsample_t, ff_cos_32768[16384]);
65 DECLARE_ALIGNED_16(fftsample_t, ff_cos_65536[32768]);
66
67 static fftsample_t *ff_cos_tabs[] = {
68 ff_cos_16, ff_cos_32, ff_cos_64, ff_cos_128, ff_cos_256,
69 ff_cos_512, ff_cos_1024, ff_cos_2048, ff_cos_4096, ff_cos_8192,
70 ff_cos_16384, ff_cos_32768, ff_cos_65536,
71 };
72
73 static int split_radix_permutation(int i, int n)
74 {
75 int m;
76 if (n <= 2)
77 return i & 1;
78 m = n >> 1;
79 if ((i & m) == 0)
80 return split_radix_permutation(i, m) * 2;
81 m >>= 1;
82 if ((i & m) == 0)
83 return split_radix_permutation(i, m) * 4 + 1;
84 else
85 return split_radix_permutation(i, m) * 4 - 1;
86 }
87
88 #define SQRTHALF (float)0.70710678118654752440 /* 1/sqrt(2) */
89
90 #define BF(x,y,a,b) {\
91 x = a - b;\
92 y = a + b;\
93 }
94
95 #define BUTTERFLIES(a0,a1,a2,a3) {\
96 BF(t3, t5, t5, t1);\
97 BF(a2.re, a0.re, a0.re, t5);\
98 BF(a3.im, a1.im, a1.im, t3);\
99 BF(t4, t6, t2, t6);\
100 BF(a3.re, a1.re, a1.re, t4);\
101 BF(a2.im, a0.im, a0.im, t6);\
102 }
103
104 // force loading all the inputs before storing any.
105 // this is slightly slower for small data, but avoids store->load aliasing
106 // for addresses separated by large powers of 2.
107 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
108 fftsample_t r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
109 BF(t3, t5, t5, t1);\
110 BF(a2.re, a0.re, r0, t5);\
111 BF(a3.im, a1.im, i1, t3);\
112 BF(t4, t6, t2, t6);\
113 BF(a3.re, a1.re, r1, t4);\
114 BF(a2.im, a0.im, i0, t6);\
115 }
116
117 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
118 t1 = a2.re * wre + a2.im * wim;\
119 t2 = a2.im * wre - a2.re * wim;\
120 t5 = a3.re * wre - a3.im * wim;\
121 t6 = a3.im * wre + a3.re * wim;\
122 BUTTERFLIES(a0,a1,a2,a3)\
123 }
124
125 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\
126 t1 = a2.re;\
127 t2 = a2.im;\
128 t5 = a3.re;\
129 t6 = a3.im;\
130 BUTTERFLIES(a0,a1,a2,a3)\
131 }
132
133 /* z[0...8n-1], w[1...2n-1] */
134 #define PASS(name)\
135 static void name(struct fft_complex *z, const fftsample_t *wre, unsigned int n)\
136 {\
137 fftsample_t t1, t2, t3, t4, t5, t6;\
138 int o1 = 2*n;\
139 int o2 = 4*n;\
140 int o3 = 6*n;\
141 const fftsample_t *wim = wre+o1;\
142 n--;\
143 \
144 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
145 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
146 do {\
147 z += 2;\
148 wre += 2;\
149 wim -= 2;\
150 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
151 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
152 } while(--n);\
153 }
154
155 PASS(pass)
156 #undef BUTTERFLIES
157 #define BUTTERFLIES BUTTERFLIES_BIG
158
159 #define DECL_FFT(n,n2,n4)\
160 static void fft##n(struct fft_complex *z)\
161 {\
162 fft##n2(z);\
163 fft##n4(z+n4*2);\
164 fft##n4(z+n4*3);\
165 pass(z,ff_cos_##n,n4/2);\
166 }
167 static void fft4(struct fft_complex *z)
168 {
169 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
170
171 BF(t3, t1, z[0].re, z[1].re);
172 BF(t8, t6, z[3].re, z[2].re);
173 BF(z[2].re, z[0].re, t1, t6);
174 BF(t4, t2, z[0].im, z[1].im);
175 BF(t7, t5, z[2].im, z[3].im);
176 BF(z[3].im, z[1].im, t4, t8);
177 BF(z[3].re, z[1].re, t3, t7);
178 BF(z[2].im, z[0].im, t2, t5);
179 }
180
181 static void fft8(struct fft_complex *z)
182 {
183 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
184
185 fft4(z);
186
187 BF(t1, z[5].re, z[4].re, -z[5].re);
188 BF(t2, z[5].im, z[4].im, -z[5].im);
189 BF(t3, z[7].re, z[6].re, -z[7].re);
190 BF(t4, z[7].im, z[6].im, -z[7].im);
191 BF(t8, t1, t3, t1);
192 BF(t7, t2, t2, t4);
193 BF(z[4].re, z[0].re, z[0].re, t1);
194 BF(z[4].im, z[0].im, z[0].im, t2);
195 BF(z[6].re, z[2].re, z[2].re, t7);
196 BF(z[6].im, z[2].im, z[2].im, t8);
197
198 TRANSFORM(z[1], z[3], z[5], z[7], SQRTHALF, SQRTHALF);
199 }
200
201 static void fft16(struct fft_complex *z)
202 {
203 fftsample_t t1, t2, t3, t4, t5, t6;
204
205 fft8(z);
206 fft4(z + 8);
207 fft4(z + 12);
208
209 TRANSFORM_ZERO(z[0], z[4], z[8], z[12]);
210 TRANSFORM(z[2], z[6], z[10], z[14], SQRTHALF, SQRTHALF);
211 TRANSFORM(z[1], z[5], z[9], z[13], ff_cos_16[1], ff_cos_16[3]);
212 TRANSFORM(z[3], z[7], z[11], z[15], ff_cos_16[3], ff_cos_16[1]);
213 }
214
215 DECL_FFT(32, 16, 8)
216 DECL_FFT(64, 32, 16)
217 DECL_FFT(128, 64, 32)
218 DECL_FFT(256, 128, 64)
219 DECL_FFT(512, 256, 128)
220
221 DECL_FFT(1024, 512, 256)
222 DECL_FFT(2048, 1024, 512)
223 DECL_FFT(4096, 2048, 1024)
224 DECL_FFT(8192, 4096, 2048)
225 DECL_FFT(16384, 8192, 4096)
226 DECL_FFT(32768, 16384, 8192)
227 DECL_FFT(65536, 32768, 16384)
228
229 static void (*fft_dispatch[]) (struct fft_complex *) = {
230 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024,
231 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536,
232 };
233
234 static void fft(struct fft_context *s, struct fft_complex *z)
235 {
236 fft_dispatch[s->nbits - 2] (z);
237 }
238
239 /* complex multiplication: p = a * b */
240 #define CMUL(pre, pim, are, aim, bre, bim) \
241 {\
242 fftsample_t _are = (are);\
243 fftsample_t _aim = (aim);\
244 fftsample_t _bre = (bre);\
245 fftsample_t _bim = (bim);\
246 (pre) = _are * _bre - _aim * _bim;\
247 (pim) = _are * _bim + _aim * _bre;\
248 }
249
250 /**
251 * Compute the middle half of the inverse MDCT of size N = 2^nbits
252 *
253 * Thus excluding the parts that can be derived by symmetry.
254 *
255 * \param output N/2 samples.
256 * \param input N/2 samples.
257 */
258 static void imdct_half(struct mdct_context *s, fftsample_t *output,
259 const fftsample_t *input)
260 {
261 int k, n8, n4, n2, n, j;
262 const uint16_t *revtab = s->fft.revtab;
263 const fftsample_t *tcos = s->tcos;
264 const fftsample_t *tsin = s->tsin;
265 const fftsample_t *in1, *in2;
266 struct fft_complex *z = (struct fft_complex *)output;
267
268 n = 1 << s->nbits;
269 n2 = n >> 1;
270 n4 = n >> 2;
271 n8 = n >> 3;
272
273 /* pre rotation */
274 in1 = input;
275 in2 = input + n2 - 1;
276 for (k = 0; k < n4; k++) {
277 j = revtab[k];
278 CMUL(z[j].re, z[j].im, *in2, *in1, tcos[k], tsin[k]);
279 in1 += 2;
280 in2 -= 2;
281 }
282 fft(&s->fft, z);
283
284 /* post rotation + reordering */
285 output += n4;
286 for (k = 0; k < n8; k++) {
287 fftsample_t r0, i0, r1, i1;
288 CMUL(r0, i1, z[n8 - k - 1].im, z[n8 - k - 1].re,
289 tsin[n8 - k - 1], tcos[n8 - k - 1]);
290 CMUL(r1, i0, z[n8 + k].im, z[n8 + k].re, tsin[n8 + k],
291 tcos[n8 + k]);
292 z[n8 - k - 1].re = r0;
293 z[n8 - k - 1].im = i0;
294 z[n8 + k].re = r1;
295 z[n8 + k].im = i1;
296 }
297 }
298
299 /**
300 * Compute the inverse MDCT of size N = 2^nbits.
301 *
302 * \param output N samples.
303 * \param input N/2 samples.
304 */
305 void imdct(struct mdct_context *s, float *output, const float *input)
306 {
307 int k;
308 int n = 1 << s->nbits;
309 int n2 = n >> 1;
310 int n4 = n >> 2;
311
312 imdct_half(s, output + n4, input);
313
314 for (k = 0; k < n4; k++) {
315 output[k] = -output[n2 - k - 1];
316 output[n - k - 1] = output[n2 + k];
317 }
318 }
319
320 static int fft_init(struct fft_context *s, int nbits)
321 {
322 int i, j, n;
323
324 if (nbits < 2 || nbits > 16)
325 return -E_FFT_BAD_PARAMS;
326 s->nbits = nbits;
327 n = 1 << nbits;
328
329 s->revtab = para_malloc(n * sizeof(uint16_t));
330 for (j = 4; j <= nbits; j++) {
331 int k = 1 << j;
332 double freq = 2 * M_PI / k;
333 fftsample_t *tab = ff_cos_tabs[j - 4];
334 for (i = 0; i <= k / 4; i++)
335 tab[i] = cos(i * freq);
336 for (i = 1; i < k / 4; i++)
337 tab[k / 2 - i] = tab[i];
338 }
339 for (i = 0; i < n; i++)
340 s->revtab[-split_radix_permutation(i, n) & (n - 1)] = i;
341 return 0;
342 }
343
344 static void fft_end(struct fft_context *ctx)
345 {
346 freep(&ctx->revtab);
347 }
348
349 /**
350 * Initialize the inverse modified cosine transform.
351 */
352 int imdct_init(int nbits, struct mdct_context **result)
353 {
354 int ret, n, n4, i;
355 double alpha;
356 struct mdct_context *s;
357
358 s = para_calloc(sizeof(*s));
359 n = 1 << nbits;
360 s->nbits = nbits;
361 s->n = n;
362 n4 = n >> 2;
363 s->tcos = para_malloc(n4 * sizeof(fftsample_t));
364 s->tsin = para_malloc(n4 * sizeof(fftsample_t));
365
366 for (i = 0; i < n4; i++) {
367 alpha = 2 * M_PI * (i + 1.0 / 8.0) / n;
368 s->tcos[i] = -cos(alpha);
369 s->tsin[i] = -sin(alpha);
370 }
371 ret = fft_init(&s->fft, s->nbits - 2);
372 if (ret < 0)
373 goto fail;
374 *result = s;
375 return 0;
376 fail:
377 freep(&s->tcos);
378 freep(&s->tsin);
379 free(s);
380 return ret;
381 }
382
383 void imdct_end(struct mdct_context *ctx)
384 {
385 freep(&ctx->tcos);
386 freep(&ctx->tsin);
387 fft_end(&ctx->fft);
388 free(ctx);
389 }