Add __aligned macro to gcc-compat.h. and use this instead of DECLARE_ALIGNED.
[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 (number of input data * 2). */
43 int n;
44 /** n = 2^n bits. */
45 int nbits;
46 /** Cosine table for pre/post rotation. */
47 fftsample_t *tcos;
48 /** Sine table for pre/post rotation. */
49 fftsample_t *tsin;
50 /** The context for the underlying fast Fourier transform. */
51 struct fft_context fft;
52 };
53
54 /** cos(2 * pi * x / n) for 0 <= x <= n / 4, followed by its reverse */
55 #define COSINE_TAB(n) fftsample_t ff_cos_ ## n[n / 2] __aligned(16)
56
57 COSINE_TAB(16);
58 COSINE_TAB(32);
59 COSINE_TAB(64);
60 COSINE_TAB(128);
61 COSINE_TAB(256);
62 COSINE_TAB(512);
63 COSINE_TAB(1024);
64 COSINE_TAB(2048);
65 COSINE_TAB(4096);
66 COSINE_TAB(8192);
67 COSINE_TAB(16384);
68 COSINE_TAB(32768);
69 COSINE_TAB(65536);
70
71 static fftsample_t *ff_cos_tabs[] = {
72 ff_cos_16, ff_cos_32, ff_cos_64, ff_cos_128, ff_cos_256,
73 ff_cos_512, ff_cos_1024, ff_cos_2048, ff_cos_4096, ff_cos_8192,
74 ff_cos_16384, ff_cos_32768, ff_cos_65536,
75 };
76
77 static int split_radix_permutation(int i, int n)
78 {
79 int m;
80 if (n <= 2)
81 return i & 1;
82 m = n >> 1;
83 if ((i & m) == 0)
84 return split_radix_permutation(i, m) * 2;
85 m >>= 1;
86 if ((i & m) == 0)
87 return split_radix_permutation(i, m) * 4 + 1;
88 else
89 return split_radix_permutation(i, m) * 4 - 1;
90 }
91
92 /** 1 / sqrt(2). */
93 #define SQRTHALF (float)0.70710678118654752440
94
95 #define BF(x,y,a,b) {\
96 x = a - b;\
97 y = a + b;\
98 }
99
100 #define BUTTERFLIES(a0,a1,a2,a3) {\
101 BF(t3, t5, t5, t1);\
102 BF(a2.re, a0.re, a0.re, t5);\
103 BF(a3.im, a1.im, a1.im, t3);\
104 BF(t4, t6, t2, t6);\
105 BF(a3.re, a1.re, a1.re, t4);\
106 BF(a2.im, a0.im, a0.im, t6);\
107 }
108
109 // force loading all the inputs before storing any.
110 // this is slightly slower for small data, but avoids store->load aliasing
111 // for addresses separated by large powers of 2.
112 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
113 fftsample_t r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
114 BF(t3, t5, t5, t1);\
115 BF(a2.re, a0.re, r0, t5);\
116 BF(a3.im, a1.im, i1, t3);\
117 BF(t4, t6, t2, t6);\
118 BF(a3.re, a1.re, r1, t4);\
119 BF(a2.im, a0.im, i0, t6);\
120 }
121
122 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
123 t1 = a2.re * wre + a2.im * wim;\
124 t2 = a2.im * wre - a2.re * wim;\
125 t5 = a3.re * wre - a3.im * wim;\
126 t6 = a3.im * wre + a3.re * wim;\
127 BUTTERFLIES(a0,a1,a2,a3)\
128 }
129
130 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\
131 t1 = a2.re;\
132 t2 = a2.im;\
133 t5 = a3.re;\
134 t6 = a3.im;\
135 BUTTERFLIES(a0,a1,a2,a3)\
136 }
137
138 /* z[0...8n-1], w[1...2n-1] */
139 #define PASS(name)\
140 static void name(struct fft_complex *z, const fftsample_t *wre, unsigned int n)\
141 {\
142 fftsample_t t1, t2, t3, t4, t5, t6;\
143 int o1 = 2*n;\
144 int o2 = 4*n;\
145 int o3 = 6*n;\
146 const fftsample_t *wim = wre+o1;\
147 n--;\
148 \
149 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
150 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
151 do {\
152 z += 2;\
153 wre += 2;\
154 wim -= 2;\
155 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
156 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
157 } while(--n);\
158 }
159
160 PASS(pass)
161 #undef BUTTERFLIES
162 #define BUTTERFLIES BUTTERFLIES_BIG
163
164 #define DECL_FFT(n,n2,n4)\
165 static void fft##n(struct fft_complex *z)\
166 {\
167 fft##n2(z);\
168 fft##n4(z+n4*2);\
169 fft##n4(z+n4*3);\
170 pass(z,ff_cos_##n,n4/2);\
171 }
172 static void fft4(struct fft_complex *z)
173 {
174 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
175
176 BF(t3, t1, z[0].re, z[1].re);
177 BF(t8, t6, z[3].re, z[2].re);
178 BF(z[2].re, z[0].re, t1, t6);
179 BF(t4, t2, z[0].im, z[1].im);
180 BF(t7, t5, z[2].im, z[3].im);
181 BF(z[3].im, z[1].im, t4, t8);
182 BF(z[3].re, z[1].re, t3, t7);
183 BF(z[2].im, z[0].im, t2, t5);
184 }
185
186 static void fft8(struct fft_complex *z)
187 {
188 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
189
190 fft4(z);
191
192 BF(t1, z[5].re, z[4].re, -z[5].re);
193 BF(t2, z[5].im, z[4].im, -z[5].im);
194 BF(t3, z[7].re, z[6].re, -z[7].re);
195 BF(t4, z[7].im, z[6].im, -z[7].im);
196 BF(t8, t1, t3, t1);
197 BF(t7, t2, t2, t4);
198 BF(z[4].re, z[0].re, z[0].re, t1);
199 BF(z[4].im, z[0].im, z[0].im, t2);
200 BF(z[6].re, z[2].re, z[2].re, t7);
201 BF(z[6].im, z[2].im, z[2].im, t8);
202
203 TRANSFORM(z[1], z[3], z[5], z[7], SQRTHALF, SQRTHALF);
204 }
205
206 static void fft16(struct fft_complex *z)
207 {
208 fftsample_t t1, t2, t3, t4, t5, t6;
209
210 fft8(z);
211 fft4(z + 8);
212 fft4(z + 12);
213
214 TRANSFORM_ZERO(z[0], z[4], z[8], z[12]);
215 TRANSFORM(z[2], z[6], z[10], z[14], SQRTHALF, SQRTHALF);
216 TRANSFORM(z[1], z[5], z[9], z[13], ff_cos_16[1], ff_cos_16[3]);
217 TRANSFORM(z[3], z[7], z[11], z[15], ff_cos_16[3], ff_cos_16[1]);
218 }
219
220 DECL_FFT(32, 16, 8)
221 DECL_FFT(64, 32, 16)
222 DECL_FFT(128, 64, 32)
223 DECL_FFT(256, 128, 64)
224 DECL_FFT(512, 256, 128)
225
226 DECL_FFT(1024, 512, 256)
227 DECL_FFT(2048, 1024, 512)
228 DECL_FFT(4096, 2048, 1024)
229 DECL_FFT(8192, 4096, 2048)
230 DECL_FFT(16384, 8192, 4096)
231 DECL_FFT(32768, 16384, 8192)
232 DECL_FFT(65536, 32768, 16384)
233
234 static void (*fft_dispatch[]) (struct fft_complex *) = {
235 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024,
236 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536,
237 };
238
239 static void fft(struct fft_context *s, struct fft_complex *z)
240 {
241 fft_dispatch[s->nbits - 2] (z);
242 }
243
244 /* complex multiplication: p = a * b */
245 #define CMUL(pre, pim, are, aim, bre, bim) \
246 {\
247 fftsample_t _are = (are);\
248 fftsample_t _aim = (aim);\
249 fftsample_t _bre = (bre);\
250 fftsample_t _bim = (bim);\
251 (pre) = _are * _bre - _aim * _bim;\
252 (pim) = _are * _bim + _aim * _bre;\
253 }
254
255 /**
256 * Compute the middle half of the inverse MDCT of size N = 2^nbits
257 *
258 * Thus excluding the parts that can be derived by symmetry.
259 *
260 * \param output N/2 samples.
261 * \param input N/2 samples.
262 */
263 static void imdct_half(struct mdct_context *s, fftsample_t *output,
264 const fftsample_t *input)
265 {
266 int k, n8, n4, n2, n, j;
267 const uint16_t *revtab = s->fft.revtab;
268 const fftsample_t *tcos = s->tcos;
269 const fftsample_t *tsin = s->tsin;
270 const fftsample_t *in1, *in2;
271 struct fft_complex *z = (struct fft_complex *)output;
272
273 n = 1 << s->nbits;
274 n2 = n >> 1;
275 n4 = n >> 2;
276 n8 = n >> 3;
277
278 /* pre rotation */
279 in1 = input;
280 in2 = input + n2 - 1;
281 for (k = 0; k < n4; k++) {
282 j = revtab[k];
283 CMUL(z[j].re, z[j].im, *in2, *in1, tcos[k], tsin[k]);
284 in1 += 2;
285 in2 -= 2;
286 }
287 fft(&s->fft, z);
288
289 /* post rotation + reordering */
290 output += n4;
291 for (k = 0; k < n8; k++) {
292 fftsample_t r0, i0, r1, i1;
293 CMUL(r0, i1, z[n8 - k - 1].im, z[n8 - k - 1].re,
294 tsin[n8 - k - 1], tcos[n8 - k - 1]);
295 CMUL(r1, i0, z[n8 + k].im, z[n8 + k].re, tsin[n8 + k],
296 tcos[n8 + k]);
297 z[n8 - k - 1].re = r0;
298 z[n8 - k - 1].im = i0;
299 z[n8 + k].re = r1;
300 z[n8 + k].im = i1;
301 }
302 }
303
304 /**
305 * Compute the inverse MDCT.
306 *
307 * \param ctx The initialized context structure.
308 * \param output N samples.
309 * \param input N/2 samples.
310 *
311 * \sa \ref imdct_init().
312 */
313 void imdct(struct mdct_context *ctx, float *output, const float *input)
314 {
315 int k;
316 int n = 1 << ctx->nbits;
317 int n2 = n >> 1;
318 int n4 = n >> 2;
319
320 imdct_half(ctx, output + n4, input);
321
322 for (k = 0; k < n4; k++) {
323 output[k] = -output[n2 - k - 1];
324 output[n - k - 1] = output[n2 + k];
325 }
326 }
327
328 static int fft_init(struct fft_context *s, int nbits)
329 {
330 int i, j, n;
331
332 if (nbits < 2 || nbits > 16)
333 return -E_FFT_BAD_PARAMS;
334 s->nbits = nbits;
335 n = 1 << nbits;
336
337 s->revtab = para_malloc(n * sizeof(uint16_t));
338 for (j = 4; j <= nbits; j++) {
339 int k = 1 << j;
340 double freq = 2 * M_PI / k;
341 fftsample_t *tab = ff_cos_tabs[j - 4];
342 for (i = 0; i <= k / 4; i++)
343 tab[i] = cos(i * freq);
344 for (i = 1; i < k / 4; i++)
345 tab[k / 2 - i] = tab[i];
346 }
347 for (i = 0; i < n; i++)
348 s->revtab[-split_radix_permutation(i, n) & (n - 1)] = i;
349 return 0;
350 }
351
352 /**
353 * Initialize the inverse modified cosine transform.
354 *
355 * \param nbits The number of bits to use (4 <= \a nbits <= 18).
356 *
357 * \param result Opaque structure that must be passed to \ref imdct().
358 *
359 * \return Standard.
360 */
361 int imdct_init(int nbits, struct mdct_context **result)
362 {
363 int ret, n, n4, i;
364 double alpha;
365 struct mdct_context *s;
366
367 s = para_calloc(sizeof(*s));
368 n = 1 << nbits;
369 s->nbits = nbits;
370 s->n = n;
371 n4 = n >> 2;
372 s->tcos = para_malloc(n4 * sizeof(fftsample_t));
373 s->tsin = para_malloc(n4 * sizeof(fftsample_t));
374
375 for (i = 0; i < n4; i++) {
376 alpha = 2 * M_PI * (i + 1.0 / 8.0) / n;
377 s->tcos[i] = -cos(alpha);
378 s->tsin[i] = -sin(alpha);
379 }
380 ret = fft_init(&s->fft, s->nbits - 2);
381 if (ret < 0)
382 goto fail;
383 *result = s;
384 return 0;
385 fail:
386 freep(&s->tcos);
387 freep(&s->tsin);
388 free(s);
389 return ret;
390 }
391
392 /**
393 * Deallocate imdct resources.
394 *
395 * \param ctx The pointer obtained by imdct_init().
396 */
397 void imdct_end(struct mdct_context *ctx)
398 {
399 free(ctx->tcos);
400 free(ctx->tsin);
401 free(ctx->fft.revtab);
402 free(ctx);
403 }