remove unused field tmp_buf of struct fft_context.
[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 };
46
47 struct mdct_context {
48 /** Size of MDCT (i.e. number of input data * 2). */
49 int n;
50 /** n = 2^n bits. */
51 int nbits;
52 /** pre/post rotation tables */
53 fftsample_t *tcos;
54 fftsample_t *tsin;
55 struct fft_context fft;
56 };
57
58 /* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */
59 DECLARE_ALIGNED_16(fftsample_t, ff_cos_16[8]);
60 DECLARE_ALIGNED_16(fftsample_t, ff_cos_32[16]);
61 DECLARE_ALIGNED_16(fftsample_t, ff_cos_64[32]);
62 DECLARE_ALIGNED_16(fftsample_t, ff_cos_128[64]);
63 DECLARE_ALIGNED_16(fftsample_t, ff_cos_256[128]);
64 DECLARE_ALIGNED_16(fftsample_t, ff_cos_512[256]);
65 DECLARE_ALIGNED_16(fftsample_t, ff_cos_1024[512]);
66 DECLARE_ALIGNED_16(fftsample_t, ff_cos_2048[1024]);
67 DECLARE_ALIGNED_16(fftsample_t, ff_cos_4096[2048]);
68 DECLARE_ALIGNED_16(fftsample_t, ff_cos_8192[4096]);
69 DECLARE_ALIGNED_16(fftsample_t, ff_cos_16384[8192]);
70 DECLARE_ALIGNED_16(fftsample_t, ff_cos_32768[16384]);
71 DECLARE_ALIGNED_16(fftsample_t, ff_cos_65536[32768]);
72
73 static fftsample_t *ff_cos_tabs[] = {
74 ff_cos_16, ff_cos_32, ff_cos_64, ff_cos_128, ff_cos_256,
75 ff_cos_512, ff_cos_1024, ff_cos_2048, ff_cos_4096, ff_cos_8192,
76 ff_cos_16384, ff_cos_32768, ff_cos_65536,
77 };
78
79 static int split_radix_permutation(int i, int n, int inverse)
80 {
81 int m;
82 if (n <= 2)
83 return i & 1;
84 m = n >> 1;
85 if (!(i & m))
86 return split_radix_permutation(i, m, inverse) * 2;
87 m >>= 1;
88 if (inverse == !(i & m))
89 return split_radix_permutation(i, m, inverse) * 4 + 1;
90 else
91 return split_radix_permutation(i, m, inverse) * 4 - 1;
92 }
93
94 #define sqrthalf (float)M_SQRT1_2
95
96 #define BF(x,y,a,b) {\
97 x = a - b;\
98 y = a + b;\
99 }
100
101 #define BUTTERFLIES(a0,a1,a2,a3) {\
102 BF(t3, t5, t5, t1);\
103 BF(a2.re, a0.re, a0.re, t5);\
104 BF(a3.im, a1.im, a1.im, t3);\
105 BF(t4, t6, t2, t6);\
106 BF(a3.re, a1.re, a1.re, t4);\
107 BF(a2.im, a0.im, a0.im, t6);\
108 }
109
110 // force loading all the inputs before storing any.
111 // this is slightly slower for small data, but avoids store->load aliasing
112 // for addresses separated by large powers of 2.
113 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
114 fftsample_t r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
115 BF(t3, t5, t5, t1);\
116 BF(a2.re, a0.re, r0, t5);\
117 BF(a3.im, a1.im, i1, t3);\
118 BF(t4, t6, t2, t6);\
119 BF(a3.re, a1.re, r1, t4);\
120 BF(a2.im, a0.im, i0, t6);\
121 }
122
123 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
124 t1 = a2.re * wre + a2.im * wim;\
125 t2 = a2.im * wre - a2.re * wim;\
126 t5 = a3.re * wre - a3.im * wim;\
127 t6 = a3.im * wre + a3.re * wim;\
128 BUTTERFLIES(a0,a1,a2,a3)\
129 }
130
131 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\
132 t1 = a2.re;\
133 t2 = a2.im;\
134 t5 = a3.re;\
135 t6 = a3.im;\
136 BUTTERFLIES(a0,a1,a2,a3)\
137 }
138
139 /* z[0...8n-1], w[1...2n-1] */
140 #define PASS(name)\
141 static void name(struct fft_complex *z, const fftsample_t *wre, unsigned int n)\
142 {\
143 fftsample_t t1, t2, t3, t4, t5, t6;\
144 int o1 = 2*n;\
145 int o2 = 4*n;\
146 int o3 = 6*n;\
147 const fftsample_t *wim = wre+o1;\
148 n--;\
149 \
150 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
151 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
152 do {\
153 z += 2;\
154 wre += 2;\
155 wim -= 2;\
156 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
157 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
158 } while(--n);\
159 }
160
161 PASS(pass)
162 #undef BUTTERFLIES
163 #define BUTTERFLIES BUTTERFLIES_BIG
164
165 #define DECL_FFT(n,n2,n4)\
166 static void fft##n(struct fft_complex *z)\
167 {\
168 fft##n2(z);\
169 fft##n4(z+n4*2);\
170 fft##n4(z+n4*3);\
171 pass(z,ff_cos_##n,n4/2);\
172 }
173 static void fft4(struct fft_complex *z)
174 {
175 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
176
177 BF(t3, t1, z[0].re, z[1].re);
178 BF(t8, t6, z[3].re, z[2].re);
179 BF(z[2].re, z[0].re, t1, t6);
180 BF(t4, t2, z[0].im, z[1].im);
181 BF(t7, t5, z[2].im, z[3].im);
182 BF(z[3].im, z[1].im, t4, t8);
183 BF(z[3].re, z[1].re, t3, t7);
184 BF(z[2].im, z[0].im, t2, t5);
185 }
186
187 static void fft8(struct fft_complex *z)
188 {
189 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
190
191 fft4(z);
192
193 BF(t1, z[5].re, z[4].re, -z[5].re);
194 BF(t2, z[5].im, z[4].im, -z[5].im);
195 BF(t3, z[7].re, z[6].re, -z[7].re);
196 BF(t4, z[7].im, z[6].im, -z[7].im);
197 BF(t8, t1, t3, t1);
198 BF(t7, t2, t2, t4);
199 BF(z[4].re, z[0].re, z[0].re, t1);
200 BF(z[4].im, z[0].im, z[0].im, t2);
201 BF(z[6].re, z[2].re, z[2].re, t7);
202 BF(z[6].im, z[2].im, z[2].im, t8);
203
204 TRANSFORM(z[1], z[3], z[5], z[7], sqrthalf, sqrthalf);
205 }
206
207 static void fft16(struct fft_complex *z)
208 {
209 fftsample_t t1, t2, t3, t4, t5, t6;
210
211 fft8(z);
212 fft4(z + 8);
213 fft4(z + 12);
214
215 TRANSFORM_ZERO(z[0], z[4], z[8], z[12]);
216 TRANSFORM(z[2], z[6], z[10], z[14], sqrthalf, sqrthalf);
217 TRANSFORM(z[1], z[5], z[9], z[13], ff_cos_16[1], ff_cos_16[3]);
218 TRANSFORM(z[3], z[7], z[11], z[15], ff_cos_16[3], ff_cos_16[1]);
219 }
220
221 DECL_FFT(32, 16, 8)
222 DECL_FFT(64, 32, 16)
223 DECL_FFT(128, 64, 32)
224 DECL_FFT(256, 128, 64)
225 DECL_FFT(512, 256, 128)
226
227 DECL_FFT(1024, 512, 256)
228 DECL_FFT(2048, 1024, 512)
229 DECL_FFT(4096, 2048, 1024)
230 DECL_FFT(8192, 4096, 2048)
231 DECL_FFT(16384, 8192, 4096)
232 DECL_FFT(32768, 16384, 8192)
233 DECL_FFT(65536, 32768, 16384)
234
235 static void (*fft_dispatch[]) (struct fft_complex *) = {
236 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024,
237 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536,
238 };
239
240 static void fft(struct fft_context *s, struct fft_complex *z)
241 {
242 fft_dispatch[s->nbits - 2] (z);
243 }
244
245 /* complex multiplication: p = a * b */
246 #define CMUL(pre, pim, are, aim, bre, bim) \
247 {\
248 fftsample_t _are = (are);\
249 fftsample_t _aim = (aim);\
250 fftsample_t _bre = (bre);\
251 fftsample_t _bim = (bim);\
252 (pre) = _are * _bre - _aim * _bim;\
253 (pim) = _are * _bim + _aim * _bre;\
254 }
255
256 /**
257 * Compute the middle half of the inverse MDCT of size N = 2^nbits
258 *
259 * Thus excluding the parts that can be derived by symmetry.
260 *
261 * \param output N/2 samples.
262 * \param input N/2 samples.
263 */
264 static void imdct_half(struct mdct_context *s, fftsample_t *output,
265 const fftsample_t *input)
266 {
267 int k, n8, n4, n2, n, j;
268 const uint16_t *revtab = s->fft.revtab;
269 const fftsample_t *tcos = s->tcos;
270 const fftsample_t *tsin = s->tsin;
271 const fftsample_t *in1, *in2;
272 struct fft_complex *z = (struct fft_complex *)output;
273
274 n = 1 << s->nbits;
275 n2 = n >> 1;
276 n4 = n >> 2;
277 n8 = n >> 3;
278
279 /* pre rotation */
280 in1 = input;
281 in2 = input + n2 - 1;
282 for (k = 0; k < n4; k++) {
283 j = revtab[k];
284 CMUL(z[j].re, z[j].im, *in2, *in1, tcos[k], tsin[k]);
285 in1 += 2;
286 in2 -= 2;
287 }
288 fft(&s->fft, z);
289
290 /* post rotation + reordering */
291 output += n4;
292 for (k = 0; k < n8; k++) {
293 fftsample_t r0, i0, r1, i1;
294 CMUL(r0, i1, z[n8 - k - 1].im, z[n8 - k - 1].re,
295 tsin[n8 - k - 1], tcos[n8 - k - 1]);
296 CMUL(r1, i0, z[n8 + k].im, z[n8 + k].re, tsin[n8 + k],
297 tcos[n8 + k]);
298 z[n8 - k - 1].re = r0;
299 z[n8 - k - 1].im = i0;
300 z[n8 + k].re = r1;
301 z[n8 + k].im = i1;
302 }
303 }
304
305 /**
306 * Compute the inverse MDCT of size N = 2^nbits.
307 *
308 * \param output N samples.
309 * \param input N/2 samples.
310 */
311 void imdct(struct mdct_context *s, float *output, const float *input)
312 {
313 int k;
314 int n = 1 << s->nbits;
315 int n2 = n >> 1;
316 int n4 = n >> 2;
317
318 imdct_half(s, output + n4, input);
319
320 for (k = 0; k < n4; k++) {
321 output[k] = -output[n2 - k - 1];
322 output[n - k - 1] = output[n2 + k];
323 }
324 }
325
326 static int fft_init(struct fft_context *s, int nbits, int inverse)
327 {
328 int i, j, n;
329
330 if (nbits < 2 || nbits > 16)
331 return -E_FFT_BAD_PARAMS;
332 s->nbits = nbits;
333 n = 1 << nbits;
334
335 s->exptab = para_malloc((n / 2) * sizeof(struct fft_complex));
336 s->revtab = para_malloc(n * sizeof(uint16_t));
337 s->inverse = inverse;
338
339 for (j = 4; j <= nbits; j++) {
340 int k = 1 << j;
341 double freq = 2 * M_PI / k;
342 fftsample_t *tab = ff_cos_tabs[j - 4];
343 for (i = 0; i <= k / 4; i++)
344 tab[i] = cos(i * freq);
345 for (i = 1; i < k / 4; i++)
346 tab[k / 2 - i] = tab[i];
347 }
348 for (i = 0; i < n; i++)
349 s->revtab[-split_radix_permutation(
350 i, n, s->inverse) & (n - 1)] = i;
351 return 0;
352 }
353
354 static void fft_end(struct fft_context *ctx)
355 {
356 freep(&ctx->revtab);
357 freep(&ctx->exptab);
358 }
359
360 DECLARE_ALIGNED(16, float, ff_sine_128[128]);
361 DECLARE_ALIGNED(16, float, ff_sine_256[256]);
362 DECLARE_ALIGNED(16, float, ff_sine_512[512]);
363 DECLARE_ALIGNED(16, float, ff_sine_1024[1024]);
364 DECLARE_ALIGNED(16, float, ff_sine_2048[2048]);
365 DECLARE_ALIGNED(16, float, ff_sine_4096[4096]);
366
367 float *ff_sine_windows[6] = {
368 ff_sine_128, ff_sine_256, ff_sine_512, ff_sine_1024,
369 ff_sine_2048, ff_sine_4096
370 };
371
372 // Generate a sine window.
373 void sine_window_init(float *window, int n)
374 {
375 int i;
376
377 for (i = 0; i < n; i++)
378 window[i] = sinf((i + 0.5) * (M_PI / (2.0 * n)));
379 }
380
381 /**
382 * Init MDCT or IMDCT computation.
383 */
384 int imdct_init(int nbits, int inverse, struct mdct_context **result)
385 {
386 int ret, n, n4, i;
387 double alpha;
388 struct mdct_context *s;
389
390 s = para_calloc(sizeof(*s));
391 n = 1 << nbits;
392 s->nbits = nbits;
393 s->n = n;
394 n4 = n >> 2;
395 s->tcos = para_malloc(n4 * sizeof(fftsample_t));
396 s->tsin = para_malloc(n4 * sizeof(fftsample_t));
397
398 for (i = 0; i < n4; i++) {
399 alpha = 2 * M_PI * (i + 1.0 / 8.0) / n;
400 s->tcos[i] = -cos(alpha);
401 s->tsin[i] = -sin(alpha);
402 }
403 ret = fft_init(&s->fft, s->nbits - 2, inverse);
404 if (ret < 0)
405 goto fail;
406 *result = s;
407 return 0;
408 fail:
409 freep(&s->tcos);
410 freep(&s->tsin);
411 free(s);
412 return ret;
413 }
414
415 void imdct_end(struct mdct_context *ctx)
416 {
417 freep(&ctx->tcos);
418 freep(&ctx->tsin);
419 fft_end(&ctx->fft);
420 free(ctx);
421 }