we always compute the inverse mdct, so get rid of the inverse bit.
[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 struct fft_complex *exptab;
40 };
41
42 struct mdct_context {
43 /** Size of MDCT (i.e. number of input data * 2). */
44 int n;
45 /** n = 2^n bits. */
46 int nbits;
47 /** pre/post rotation tables */
48 fftsample_t *tcos;
49 fftsample_t *tsin;
50 struct fft_context fft;
51 };
52
53 /* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */
54 DECLARE_ALIGNED_16(fftsample_t, ff_cos_16[8]);
55 DECLARE_ALIGNED_16(fftsample_t, ff_cos_32[16]);
56 DECLARE_ALIGNED_16(fftsample_t, ff_cos_64[32]);
57 DECLARE_ALIGNED_16(fftsample_t, ff_cos_128[64]);
58 DECLARE_ALIGNED_16(fftsample_t, ff_cos_256[128]);
59 DECLARE_ALIGNED_16(fftsample_t, ff_cos_512[256]);
60 DECLARE_ALIGNED_16(fftsample_t, ff_cos_1024[512]);
61 DECLARE_ALIGNED_16(fftsample_t, ff_cos_2048[1024]);
62 DECLARE_ALIGNED_16(fftsample_t, ff_cos_4096[2048]);
63 DECLARE_ALIGNED_16(fftsample_t, ff_cos_8192[4096]);
64 DECLARE_ALIGNED_16(fftsample_t, ff_cos_16384[8192]);
65 DECLARE_ALIGNED_16(fftsample_t, ff_cos_32768[16384]);
66 DECLARE_ALIGNED_16(fftsample_t, ff_cos_65536[32768]);
67
68 static fftsample_t *ff_cos_tabs[] = {
69 ff_cos_16, ff_cos_32, ff_cos_64, ff_cos_128, ff_cos_256,
70 ff_cos_512, ff_cos_1024, ff_cos_2048, ff_cos_4096, ff_cos_8192,
71 ff_cos_16384, ff_cos_32768, ff_cos_65536,
72 };
73
74 static int split_radix_permutation(int i, int n)
75 {
76 int m;
77 if (n <= 2)
78 return i & 1;
79 m = n >> 1;
80 if ((i & m) == 0)
81 return split_radix_permutation(i, m) * 2;
82 m >>= 1;
83 if ((i & m) == 0)
84 return split_radix_permutation(i, m) * 4 + 1;
85 else
86 return split_radix_permutation(i, m) * 4 - 1;
87 }
88
89 #define SQRTHALF (float)0.70710678118654752440 /* 1/sqrt(2) */
90
91 #define BF(x,y,a,b) {\
92 x = a - b;\
93 y = a + b;\
94 }
95
96 #define BUTTERFLIES(a0,a1,a2,a3) {\
97 BF(t3, t5, t5, t1);\
98 BF(a2.re, a0.re, a0.re, t5);\
99 BF(a3.im, a1.im, a1.im, t3);\
100 BF(t4, t6, t2, t6);\
101 BF(a3.re, a1.re, a1.re, t4);\
102 BF(a2.im, a0.im, a0.im, t6);\
103 }
104
105 // force loading all the inputs before storing any.
106 // this is slightly slower for small data, but avoids store->load aliasing
107 // for addresses separated by large powers of 2.
108 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
109 fftsample_t r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
110 BF(t3, t5, t5, t1);\
111 BF(a2.re, a0.re, r0, t5);\
112 BF(a3.im, a1.im, i1, t3);\
113 BF(t4, t6, t2, t6);\
114 BF(a3.re, a1.re, r1, t4);\
115 BF(a2.im, a0.im, i0, t6);\
116 }
117
118 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
119 t1 = a2.re * wre + a2.im * wim;\
120 t2 = a2.im * wre - a2.re * wim;\
121 t5 = a3.re * wre - a3.im * wim;\
122 t6 = a3.im * wre + a3.re * wim;\
123 BUTTERFLIES(a0,a1,a2,a3)\
124 }
125
126 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\
127 t1 = a2.re;\
128 t2 = a2.im;\
129 t5 = a3.re;\
130 t6 = a3.im;\
131 BUTTERFLIES(a0,a1,a2,a3)\
132 }
133
134 /* z[0...8n-1], w[1...2n-1] */
135 #define PASS(name)\
136 static void name(struct fft_complex *z, const fftsample_t *wre, unsigned int n)\
137 {\
138 fftsample_t t1, t2, t3, t4, t5, t6;\
139 int o1 = 2*n;\
140 int o2 = 4*n;\
141 int o3 = 6*n;\
142 const fftsample_t *wim = wre+o1;\
143 n--;\
144 \
145 TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
146 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
147 do {\
148 z += 2;\
149 wre += 2;\
150 wim -= 2;\
151 TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
152 TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
153 } while(--n);\
154 }
155
156 PASS(pass)
157 #undef BUTTERFLIES
158 #define BUTTERFLIES BUTTERFLIES_BIG
159
160 #define DECL_FFT(n,n2,n4)\
161 static void fft##n(struct fft_complex *z)\
162 {\
163 fft##n2(z);\
164 fft##n4(z+n4*2);\
165 fft##n4(z+n4*3);\
166 pass(z,ff_cos_##n,n4/2);\
167 }
168 static void fft4(struct fft_complex *z)
169 {
170 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
171
172 BF(t3, t1, z[0].re, z[1].re);
173 BF(t8, t6, z[3].re, z[2].re);
174 BF(z[2].re, z[0].re, t1, t6);
175 BF(t4, t2, z[0].im, z[1].im);
176 BF(t7, t5, z[2].im, z[3].im);
177 BF(z[3].im, z[1].im, t4, t8);
178 BF(z[3].re, z[1].re, t3, t7);
179 BF(z[2].im, z[0].im, t2, t5);
180 }
181
182 static void fft8(struct fft_complex *z)
183 {
184 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
185
186 fft4(z);
187
188 BF(t1, z[5].re, z[4].re, -z[5].re);
189 BF(t2, z[5].im, z[4].im, -z[5].im);
190 BF(t3, z[7].re, z[6].re, -z[7].re);
191 BF(t4, z[7].im, z[6].im, -z[7].im);
192 BF(t8, t1, t3, t1);
193 BF(t7, t2, t2, t4);
194 BF(z[4].re, z[0].re, z[0].re, t1);
195 BF(z[4].im, z[0].im, z[0].im, t2);
196 BF(z[6].re, z[2].re, z[2].re, t7);
197 BF(z[6].im, z[2].im, z[2].im, t8);
198
199 TRANSFORM(z[1], z[3], z[5], z[7], SQRTHALF, SQRTHALF);
200 }
201
202 static void fft16(struct fft_complex *z)
203 {
204 fftsample_t t1, t2, t3, t4, t5, t6;
205
206 fft8(z);
207 fft4(z + 8);
208 fft4(z + 12);
209
210 TRANSFORM_ZERO(z[0], z[4], z[8], z[12]);
211 TRANSFORM(z[2], z[6], z[10], z[14], SQRTHALF, SQRTHALF);
212 TRANSFORM(z[1], z[5], z[9], z[13], ff_cos_16[1], ff_cos_16[3]);
213 TRANSFORM(z[3], z[7], z[11], z[15], ff_cos_16[3], ff_cos_16[1]);
214 }
215
216 DECL_FFT(32, 16, 8)
217 DECL_FFT(64, 32, 16)
218 DECL_FFT(128, 64, 32)
219 DECL_FFT(256, 128, 64)
220 DECL_FFT(512, 256, 128)
221
222 DECL_FFT(1024, 512, 256)
223 DECL_FFT(2048, 1024, 512)
224 DECL_FFT(4096, 2048, 1024)
225 DECL_FFT(8192, 4096, 2048)
226 DECL_FFT(16384, 8192, 4096)
227 DECL_FFT(32768, 16384, 8192)
228 DECL_FFT(65536, 32768, 16384)
229
230 static void (*fft_dispatch[]) (struct fft_complex *) = {
231 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024,
232 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536,
233 };
234
235 static void fft(struct fft_context *s, struct fft_complex *z)
236 {
237 fft_dispatch[s->nbits - 2] (z);
238 }
239
240 /* complex multiplication: p = a * b */
241 #define CMUL(pre, pim, are, aim, bre, bim) \
242 {\
243 fftsample_t _are = (are);\
244 fftsample_t _aim = (aim);\
245 fftsample_t _bre = (bre);\
246 fftsample_t _bim = (bim);\
247 (pre) = _are * _bre - _aim * _bim;\
248 (pim) = _are * _bim + _aim * _bre;\
249 }
250
251 /**
252 * Compute the middle half of the inverse MDCT of size N = 2^nbits
253 *
254 * Thus excluding the parts that can be derived by symmetry.
255 *
256 * \param output N/2 samples.
257 * \param input N/2 samples.
258 */
259 static void imdct_half(struct mdct_context *s, fftsample_t *output,
260 const fftsample_t *input)
261 {
262 int k, n8, n4, n2, n, j;
263 const uint16_t *revtab = s->fft.revtab;
264 const fftsample_t *tcos = s->tcos;
265 const fftsample_t *tsin = s->tsin;
266 const fftsample_t *in1, *in2;
267 struct fft_complex *z = (struct fft_complex *)output;
268
269 n = 1 << s->nbits;
270 n2 = n >> 1;
271 n4 = n >> 2;
272 n8 = n >> 3;
273
274 /* pre rotation */
275 in1 = input;
276 in2 = input + n2 - 1;
277 for (k = 0; k < n4; k++) {
278 j = revtab[k];
279 CMUL(z[j].re, z[j].im, *in2, *in1, tcos[k], tsin[k]);
280 in1 += 2;
281 in2 -= 2;
282 }
283 fft(&s->fft, z);
284
285 /* post rotation + reordering */
286 output += n4;
287 for (k = 0; k < n8; k++) {
288 fftsample_t r0, i0, r1, i1;
289 CMUL(r0, i1, z[n8 - k - 1].im, z[n8 - k - 1].re,
290 tsin[n8 - k - 1], tcos[n8 - k - 1]);
291 CMUL(r1, i0, z[n8 + k].im, z[n8 + k].re, tsin[n8 + k],
292 tcos[n8 + k]);
293 z[n8 - k - 1].re = r0;
294 z[n8 - k - 1].im = i0;
295 z[n8 + k].re = r1;
296 z[n8 + k].im = i1;
297 }
298 }
299
300 /**
301 * Compute the inverse MDCT of size N = 2^nbits.
302 *
303 * \param output N samples.
304 * \param input N/2 samples.
305 */
306 void imdct(struct mdct_context *s, float *output, const float *input)
307 {
308 int k;
309 int n = 1 << s->nbits;
310 int n2 = n >> 1;
311 int n4 = n >> 2;
312
313 imdct_half(s, output + n4, input);
314
315 for (k = 0; k < n4; k++) {
316 output[k] = -output[n2 - k - 1];
317 output[n - k - 1] = output[n2 + k];
318 }
319 }
320
321 static int fft_init(struct fft_context *s, int nbits)
322 {
323 int i, j, n;
324
325 if (nbits < 2 || nbits > 16)
326 return -E_FFT_BAD_PARAMS;
327 s->nbits = nbits;
328 n = 1 << nbits;
329
330 s->exptab = para_malloc((n / 2) * sizeof(struct fft_complex));
331 s->revtab = para_malloc(n * sizeof(uint16_t));
332 for (j = 4; j <= nbits; j++) {
333 int k = 1 << j;
334 double freq = 2 * M_PI / k;
335 fftsample_t *tab = ff_cos_tabs[j - 4];
336 for (i = 0; i <= k / 4; i++)
337 tab[i] = cos(i * freq);
338 for (i = 1; i < k / 4; i++)
339 tab[k / 2 - i] = tab[i];
340 }
341 for (i = 0; i < n; i++)
342 s->revtab[-split_radix_permutation(i, n) & (n - 1)] = i;
343 return 0;
344 }
345
346 static void fft_end(struct fft_context *ctx)
347 {
348 freep(&ctx->revtab);
349 freep(&ctx->exptab);
350 }
351
352 /**
353 * Initialize the inverse modified cosine transform.
354 */
355 int imdct_init(int nbits, struct mdct_context **result)
356 {
357 int ret, n, n4, i;
358 double alpha;
359 struct mdct_context *s;
360
361 s = para_calloc(sizeof(*s));
362 n = 1 << nbits;
363 s->nbits = nbits;
364 s->n = n;
365 n4 = n >> 2;
366 s->tcos = para_malloc(n4 * sizeof(fftsample_t));
367 s->tsin = para_malloc(n4 * sizeof(fftsample_t));
368
369 for (i = 0; i < n4; i++) {
370 alpha = 2 * M_PI * (i + 1.0 / 8.0) / n;
371 s->tcos[i] = -cos(alpha);
372 s->tsin[i] = -sin(alpha);
373 }
374 ret = fft_init(&s->fft, s->nbits - 2);
375 if (ret < 0)
376 goto fail;
377 *result = s;
378 return 0;
379 fail:
380 freep(&s->tcos);
381 freep(&s->tsin);
382 free(s);
383 return ret;
384 }
385
386 void imdct_end(struct mdct_context *ctx)
387 {
388 freep(&ctx->tcos);
389 freep(&ctx->tsin);
390 fft_end(&ctx->fft);
391 free(ctx);
392 }