simplify fft_init() due to split_radix is always 1.
[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 #define DECLARE_ALIGNED(n,t,v) t v __attribute__ ((aligned (n)))
33 #define DECLARE_ALIGNED_16(t, v) DECLARE_ALIGNED(16, t, v)
34 #define M_SQRT1_2 0.70710678118654752440 /* 1/sqrt(2) */
35
36 struct fft_complex {
37 fftsample_t re, im;
38 };
39
40 struct fft_context {
41 int nbits;
42 int inverse;
43 uint16_t *revtab;
44 struct fft_complex *exptab;
45 struct fft_complex *exptab1; /* only used by SSE code */
46 struct fft_complex *tmp_buf;
47 };
48
49 struct mdct_context {
50 /** Size of MDCT (i.e. number of input data * 2). */
51 int n;
52 /** n = 2^n bits. */
53 int nbits;
54 /** pre/post rotation tables */
55 fftsample_t *tcos;
56 fftsample_t *tsin;
57 struct fft_context fft;
58 };
59
60 /* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */
61 DECLARE_ALIGNED_16(fftsample_t, ff_cos_16[8]);
62 DECLARE_ALIGNED_16(fftsample_t, ff_cos_32[16]);
63 DECLARE_ALIGNED_16(fftsample_t, ff_cos_64[32]);
64 DECLARE_ALIGNED_16(fftsample_t, ff_cos_128[64]);
65 DECLARE_ALIGNED_16(fftsample_t, ff_cos_256[128]);
66 DECLARE_ALIGNED_16(fftsample_t, ff_cos_512[256]);
67 DECLARE_ALIGNED_16(fftsample_t, ff_cos_1024[512]);
68 DECLARE_ALIGNED_16(fftsample_t, ff_cos_2048[1024]);
69 DECLARE_ALIGNED_16(fftsample_t, ff_cos_4096[2048]);
70 DECLARE_ALIGNED_16(fftsample_t, ff_cos_8192[4096]);
71 DECLARE_ALIGNED_16(fftsample_t, ff_cos_16384[8192]);
72 DECLARE_ALIGNED_16(fftsample_t, ff_cos_32768[16384]);
73 DECLARE_ALIGNED_16(fftsample_t, ff_cos_65536[32768]);
74
75 static fftsample_t *ff_cos_tabs[] = {
76 ff_cos_16, ff_cos_32, ff_cos_64, ff_cos_128, ff_cos_256,
77 ff_cos_512, ff_cos_1024, ff_cos_2048, ff_cos_4096, ff_cos_8192,
78 ff_cos_16384, ff_cos_32768, ff_cos_65536,
79 };
80
81 static int split_radix_permutation(int i, int n, int inverse)
82 {
83 int m;
84 if (n <= 2)
85 return i & 1;
86 m = n >> 1;
87 if (!(i & m))
88 return split_radix_permutation(i, m, inverse) * 2;
89 m >>= 1;
90 if (inverse == !(i & m))
91 return split_radix_permutation(i, m, inverse) * 4 + 1;
92 else
93 return split_radix_permutation(i, m, inverse) * 4 - 1;
94 }
95
96 #define sqrthalf (float)M_SQRT1_2
97
98 #define BF(x,y,a,b) {\
99 x = a - b;\
100 y = a + b;\
101 }
102
103 #define BUTTERFLIES(a0,a1,a2,a3) {\
104 BF(t3, t5, t5, t1);\
105 BF(a2.re, a0.re, a0.re, t5);\
106 BF(a3.im, a1.im, a1.im, t3);\
107 BF(t4, t6, t2, t6);\
108 BF(a3.re, a1.re, a1.re, t4);\
109 BF(a2.im, a0.im, a0.im, t6);\
110 }
111
112 // force loading all the inputs before storing any.
113 // this is slightly slower for small data, but avoids store->load aliasing
114 // for addresses separated by large powers of 2.
115 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
116 fftsample_t r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
117 BF(t3, t5, t5, t1);\
118 BF(a2.re, a0.re, r0, t5);\
119 BF(a3.im, a1.im, i1, t3);\
120 BF(t4, t6, t2, t6);\
121 BF(a3.re, a1.re, r1, t4);\
122 BF(a2.im, a0.im, i0, t6);\
123 }
124
125 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
126 t1 = a2.re * wre + a2.im * wim;\
127 t2 = a2.im * wre - a2.re * wim;\
128 t5 = a3.re * wre - a3.im * wim;\
129 t6 = a3.im * wre + a3.re * wim;\
130 BUTTERFLIES(a0,a1,a2,a3)\
131 }
132
133 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\
134 t1 = a2.re;\
135 t2 = a2.im;\
136 t5 = a3.re;\
137 t6 = a3.im;\
138 BUTTERFLIES(a0,a1,a2,a3)\
139 }
140
141 /* z[0...8n-1], w[1...2n-1] */
142 #define PASS(name)\
143 static void name(struct fft_complex *z, const fftsample_t *wre, unsigned int n)\
144 {\
145 fftsample_t t1, t2, t3, t4, t5, t6;\
146 int o1 = 2*n;\
147 int o2 = 4*n;\
148 int o3 = 6*n;\
149 const fftsample_t *wim = wre+o1;\
150 n--;\
151 \
152 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
153 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
154 do {\
155 z += 2;\
156 wre += 2;\
157 wim -= 2;\
158 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
159 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
160 } while(--n);\
161 }
162
163 PASS(pass)
164 #undef BUTTERFLIES
165 #define BUTTERFLIES BUTTERFLIES_BIG
166
167 #define DECL_FFT(n,n2,n4)\
168 static void fft##n(struct fft_complex *z)\
169 {\
170 fft##n2(z);\
171 fft##n4(z+n4*2);\
172 fft##n4(z+n4*3);\
173 pass(z,ff_cos_##n,n4/2);\
174 }
175 static void fft4(struct fft_complex *z)
176 {
177 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
178
179 BF(t3, t1, z[0].re, z[1].re);
180 BF(t8, t6, z[3].re, z[2].re);
181 BF(z[2].re, z[0].re, t1, t6);
182 BF(t4, t2, z[0].im, z[1].im);
183 BF(t7, t5, z[2].im, z[3].im);
184 BF(z[3].im, z[1].im, t4, t8);
185 BF(z[3].re, z[1].re, t3, t7);
186 BF(z[2].im, z[0].im, t2, t5);
187 }
188
189 static void fft8(struct fft_complex *z)
190 {
191 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
192
193 fft4(z);
194
195 BF(t1, z[5].re, z[4].re, -z[5].re);
196 BF(t2, z[5].im, z[4].im, -z[5].im);
197 BF(t3, z[7].re, z[6].re, -z[7].re);
198 BF(t4, z[7].im, z[6].im, -z[7].im);
199 BF(t8, t1, t3, t1);
200 BF(t7, t2, t2, t4);
201 BF(z[4].re, z[0].re, z[0].re, t1);
202 BF(z[4].im, z[0].im, z[0].im, t2);
203 BF(z[6].re, z[2].re, z[2].re, t7);
204 BF(z[6].im, z[2].im, z[2].im, t8);
205
206 TRANSFORM(z[1], z[3], z[5], z[7], sqrthalf, sqrthalf);
207 }
208
209 static void fft16(struct fft_complex *z)
210 {
211 fftsample_t t1, t2, t3, t4, t5, t6;
212
213 fft8(z);
214 fft4(z + 8);
215 fft4(z + 12);
216
217 TRANSFORM_ZERO(z[0], z[4], z[8], z[12]);
218 TRANSFORM(z[2], z[6], z[10], z[14], sqrthalf, sqrthalf);
219 TRANSFORM(z[1], z[5], z[9], z[13], ff_cos_16[1], ff_cos_16[3]);
220 TRANSFORM(z[3], z[7], z[11], z[15], ff_cos_16[3], ff_cos_16[1]);
221 }
222
223 DECL_FFT(32, 16, 8)
224 DECL_FFT(64, 32, 16)
225 DECL_FFT(128, 64, 32)
226 DECL_FFT(256, 128, 64)
227 DECL_FFT(512, 256, 128)
228
229 DECL_FFT(1024, 512, 256)
230 DECL_FFT(2048, 1024, 512)
231 DECL_FFT(4096, 2048, 1024)
232 DECL_FFT(8192, 4096, 2048)
233 DECL_FFT(16384, 8192, 4096)
234 DECL_FFT(32768, 16384, 8192)
235 DECL_FFT(65536, 32768, 16384)
236
237 static void (*fft_dispatch[]) (struct fft_complex *) = {
238 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024,
239 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536,
240 };
241
242 static void fft(struct fft_context *s, struct fft_complex *z)
243 {
244 fft_dispatch[s->nbits - 2] (z);
245 }
246
247 /* complex multiplication: p = a * b */
248 #define CMUL(pre, pim, are, aim, bre, bim) \
249 {\
250 fftsample_t _are = (are);\
251 fftsample_t _aim = (aim);\
252 fftsample_t _bre = (bre);\
253 fftsample_t _bim = (bim);\
254 (pre) = _are * _bre - _aim * _bim;\
255 (pim) = _are * _bim + _aim * _bre;\
256 }
257
258 /**
259 * Compute the middle half of the inverse MDCT of size N = 2^nbits
260 *
261 * Thus excluding the parts that can be derived by symmetry.
262 *
263 * \param output N/2 samples.
264 * \param input N/2 samples.
265 */
266 static void imdct_half(struct mdct_context *s, fftsample_t *output,
267 const fftsample_t *input)
268 {
269 int k, n8, n4, n2, n, j;
270 const uint16_t *revtab = s->fft.revtab;
271 const fftsample_t *tcos = s->tcos;
272 const fftsample_t *tsin = s->tsin;
273 const fftsample_t *in1, *in2;
274 struct fft_complex *z = (struct fft_complex *)output;
275
276 n = 1 << s->nbits;
277 n2 = n >> 1;
278 n4 = n >> 2;
279 n8 = n >> 3;
280
281 /* pre rotation */
282 in1 = input;
283 in2 = input + n2 - 1;
284 for (k = 0; k < n4; k++) {
285 j = revtab[k];
286 CMUL(z[j].re, z[j].im, *in2, *in1, tcos[k], tsin[k]);
287 in1 += 2;
288 in2 -= 2;
289 }
290 fft(&s->fft, z);
291
292 /* post rotation + reordering */
293 output += n4;
294 for (k = 0; k < n8; k++) {
295 fftsample_t r0, i0, r1, i1;
296 CMUL(r0, i1, z[n8 - k - 1].im, z[n8 - k - 1].re,
297 tsin[n8 - k - 1], tcos[n8 - k - 1]);
298 CMUL(r1, i0, z[n8 + k].im, z[n8 + k].re, tsin[n8 + k],
299 tcos[n8 + k]);
300 z[n8 - k - 1].re = r0;
301 z[n8 - k - 1].im = i0;
302 z[n8 + k].re = r1;
303 z[n8 + k].im = i1;
304 }
305 }
306
307 /**
308 * Compute the inverse MDCT of size N = 2^nbits.
309 *
310 * \param output N samples.
311 * \param input N/2 samples.
312 */
313 void imdct(struct mdct_context *s, float *output, const float *input)
314 {
315 int k;
316 int n = 1 << s->nbits;
317 int n2 = n >> 1;
318 int n4 = n >> 2;
319
320 imdct_half(s, 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, int inverse)
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->tmp_buf = NULL;
338 s->exptab = para_malloc((n / 2) * sizeof(struct fft_complex));
339 s->revtab = para_malloc(n * sizeof(uint16_t));
340 s->inverse = inverse;
341
342 s->exptab1 = NULL;
343
344 for (j = 4; j <= nbits; j++) {
345 int k = 1 << j;
346 double freq = 2 * M_PI / k;
347 fftsample_t *tab = ff_cos_tabs[j - 4];
348 for (i = 0; i <= k / 4; i++)
349 tab[i] = cos(i * freq);
350 for (i = 1; i < k / 4; i++)
351 tab[k / 2 - i] = tab[i];
352 }
353 for (i = 0; i < n; i++)
354 s->revtab[-split_radix_permutation(
355 i, n, s->inverse) & (n - 1)] = i;
356 s->tmp_buf = para_malloc(n * sizeof(struct fft_complex));
357 return 0;
358 }
359
360 static void fft_end(struct fft_context *ctx)
361 {
362 freep(&ctx->revtab);
363 freep(&ctx->exptab);
364 freep(&ctx->exptab1);
365 freep(&ctx->tmp_buf);
366 }
367
368 DECLARE_ALIGNED(16, float, ff_sine_128[128]);
369 DECLARE_ALIGNED(16, float, ff_sine_256[256]);
370 DECLARE_ALIGNED(16, float, ff_sine_512[512]);
371 DECLARE_ALIGNED(16, float, ff_sine_1024[1024]);
372 DECLARE_ALIGNED(16, float, ff_sine_2048[2048]);
373 DECLARE_ALIGNED(16, float, ff_sine_4096[4096]);
374
375 float *ff_sine_windows[6] = {
376 ff_sine_128, ff_sine_256, ff_sine_512, ff_sine_1024,
377 ff_sine_2048, ff_sine_4096
378 };
379
380 // Generate a sine window.
381 void sine_window_init(float *window, int n)
382 {
383 int i;
384
385 for (i = 0; i < n; i++)
386 window[i] = sinf((i + 0.5) * (M_PI / (2.0 * n)));
387 }
388
389 /**
390 * Init MDCT or IMDCT computation.
391 */
392 int imdct_init(int nbits, int inverse, struct mdct_context **result)
393 {
394 int ret, n, n4, i;
395 double alpha;
396 struct mdct_context *s;
397
398 s = para_calloc(sizeof(*s));
399 n = 1 << nbits;
400 s->nbits = nbits;
401 s->n = n;
402 n4 = n >> 2;
403 s->tcos = para_malloc(n4 * sizeof(fftsample_t));
404 s->tsin = para_malloc(n4 * sizeof(fftsample_t));
405
406 for (i = 0; i < n4; i++) {
407 alpha = 2 * M_PI * (i + 1.0 / 8.0) / n;
408 s->tcos[i] = -cos(alpha);
409 s->tsin[i] = -sin(alpha);
410 }
411 ret = fft_init(&s->fft, s->nbits - 2, inverse);
412 if (ret < 0)
413 goto fail;
414 *result = s;
415 return 0;
416 fail:
417 freep(&s->tcos);
418 freep(&s->tsin);
419 free(s);
420 return ret;
421 }
422
423 void imdct_end(struct mdct_context *ctx)
424 {
425 freep(&ctx->tcos);
426 freep(&ctx->tsin);
427 fft_end(&ctx->fft);
428 free(ctx);
429 }