4 * Extracted 2009 from mplayer 2009-02-10 libavcodec/fft.c and libavcodec/mdct.c
6 * Copyright (c) 2008 Loren Merritt
7 * Copyright (c) 2002 Fabrice Bellard
8 * Partly based on libdjbfft by D. J. Bernstein
10 * Licensed under the GNU Lesser General Public License.
11 * For licencing details see COPYING.LIB.
15 * \file imdct.c Inverse modified discrete cosine transform.
30 typedef float fftsample_t;
32 /** Canonical representation of a complex number. */
36 /** Imaginary part. */
40 /** FFT Lookup table. */
42 /** Number of bits of this instance of the FFT. */
44 /** The lookup table for cosine values. */
49 /** Size of MDCT (number of input data * 2). */
53 /** Cosine table for pre/post rotation. */
55 /** Sine table for pre/post rotation. */
57 /** The context for the underlying fast Fourier transform. */
58 struct fft_context fft;
61 /** cos(2 * pi * x / n) for 0 <= x <= n / 4, followed by its reverse */
62 #define COSINE_TAB(n) static fftsample_t cos_ ## n[n / 2] __a_aligned(16)
78 static fftsample_t *cos_tabs[] = {
79 cos_16, cos_32, cos_64, cos_128, cos_256, cos_512, cos_1024, cos_2048,
80 cos_4096, cos_8192, cos_16384, cos_32768, cos_65536,
83 __a_const static int split_radix_permutation(int i, int n)
90 return split_radix_permutation(i, m) * 2;
93 return split_radix_permutation(i, m) * 4 + 1;
95 return split_radix_permutation(i, m) * 4 - 1;
98 #define BF(x, y, a, b) {\
103 #define BUTTERFLIES(a0, a1, a2, a3) {\
105 BF(a2.re, a0.re, a0.re, t5);\
106 BF(a3.im, a1.im, a1.im, t3);\
108 BF(a3.re, a1.re, a1.re, t4);\
109 BF(a2.im, a0.im, a0.im, t6);\
113 * Force loading all the inputs before storing any. This is slightly slower for
114 * small data, but avoids store->load aliasing for addresses separated by large
117 #define BUTTERFLIES_BIG(a0, a1, a2, a3) {\
118 fftsample_t r0 = a0.re, i0 = a0.im, r1 = a1.re, i1 = a1.im;\
120 BF(a2.re, a0.re, r0, t5);\
121 BF(a3.im, a1.im, i1, t3);\
123 BF(a3.re, a1.re, r1, t4);\
124 BF(a2.im, a0.im, i0, t6);\
127 #define TRANSFORM(a0, a1, a2, a3, wre,wim) {\
128 t1 = a2.re * wre + a2.im * wim;\
129 t2 = a2.im * wre - a2.re * wim;\
130 t5 = a3.re * wre - a3.im * wim;\
131 t6 = a3.im * wre + a3.re * wim;\
132 BUTTERFLIES(a0, a1, a2, a3)\
135 #define TRANSFORM_ZERO(a0, a1, a2, a3) {\
140 BUTTERFLIES(a0, a1, a2, a3)\
143 /* z[0...8n - 1], w[1...2n - 1] */
145 static void name(struct fft_complex *z, const fftsample_t *wre, unsigned int n)\
147 fftsample_t t1, t2, t3, t4, t5, t6;\
151 const fftsample_t *wim = wre + o1;\
154 TRANSFORM_ZERO(z[0], z[o1], z[o2], z[o3]);\
155 TRANSFORM(z[1], z[o1 + 1], z[o2 + 1], z[o3 + 1], wre[1], wim[-1]);\
160 TRANSFORM(z[0], z[o1], z[o2], z[o3], wre[0], wim[0]);\
161 TRANSFORM(z[1], z[o1 + 1], z[o2 + 1], z[o3 + 1], wre[1], wim[-1]);\
167 #define BUTTERFLIES BUTTERFLIES_BIG
169 #define DECL_FFT(n, n2, n4)\
170 static void fft##n(struct fft_complex *z)\
173 fft ## n4(z + n4 * 2);\
174 fft ## n4(z + n4 * 3);\
175 pass(z, cos_ ## n, n4 / 2);\
178 static void fft4(struct fft_complex *z)
180 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
182 BF(t3, t1, z[0].re, z[1].re);
183 BF(t8, t6, z[3].re, z[2].re);
184 BF(z[2].re, z[0].re, t1, t6);
185 BF(t4, t2, z[0].im, z[1].im);
186 BF(t7, t5, z[2].im, z[3].im);
187 BF(z[3].im, z[1].im, t4, t8);
188 BF(z[3].re, z[1].re, t3, t7);
189 BF(z[2].im, z[0].im, t2, t5);
192 static void fft8(struct fft_complex *z)
194 fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
198 BF(t1, z[5].re, z[4].re, -z[5].re);
199 BF(t2, z[5].im, z[4].im, -z[5].im);
200 BF(t3, z[7].re, z[6].re, -z[7].re);
201 BF(t4, z[7].im, z[6].im, -z[7].im);
204 BF(z[4].re, z[0].re, z[0].re, t1);
205 BF(z[4].im, z[0].im, z[0].im, t2);
206 BF(z[6].re, z[2].re, z[2].re, t7);
207 BF(z[6].im, z[2].im, z[2].im, t8);
209 TRANSFORM(z[1], z[3], z[5], z[7], M_SQRT1_2, M_SQRT1_2);
212 static void fft16(struct fft_complex *z)
214 fftsample_t t1, t2, t3, t4, t5, t6;
220 TRANSFORM_ZERO(z[0], z[4], z[8], z[12]);
221 TRANSFORM(z[2], z[6], z[10], z[14], M_SQRT1_2, M_SQRT1_2);
222 TRANSFORM(z[1], z[5], z[9], z[13], cos_16[1], cos_16[3]);
223 TRANSFORM(z[3], z[7], z[11], z[15], cos_16[3], cos_16[1]);
228 DECL_FFT(128, 64, 32)
229 DECL_FFT(256, 128, 64)
230 DECL_FFT(512, 256, 128)
232 DECL_FFT(1024, 512, 256)
233 DECL_FFT(2048, 1024, 512)
234 DECL_FFT(4096, 2048, 1024)
235 DECL_FFT(8192, 4096, 2048)
236 DECL_FFT(16384, 8192, 4096)
237 DECL_FFT(32768, 16384, 8192)
238 DECL_FFT(65536, 32768, 16384)
240 static void (*fft_dispatch[]) (struct fft_complex *) = {
241 fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512, fft1024,
242 fft2048, fft4096, fft8192, fft16384, fft32768, fft65536,
245 static void fft(struct fft_context *s, struct fft_complex *z)
247 fft_dispatch[s->nbits - 2] (z);
250 /* complex multiplication: p = a * b */
251 #define CMUL(pre, pim, are, aim, bre, bim) \
253 fftsample_t _are = (are);\
254 fftsample_t _aim = (aim);\
255 fftsample_t _bre = (bre);\
256 fftsample_t _bim = (bim);\
257 (pre) = _are * _bre - _aim * _bim;\
258 (pim) = _are * _bim + _aim * _bre;\
262 * Compute the middle half of the inverse MDCT of size N = 2^nbits
264 * Thus excluding the parts that can be derived by symmetry.
266 * \param output N/2 samples.
267 * \param input N/2 samples.
269 static void imdct_half(struct mdct_context *s, fftsample_t *output,
270 const fftsample_t *input)
272 int k, n8, n4, n2, n, j;
273 const uint16_t *revtab = s->fft.revtab;
274 const fftsample_t *tcos = s->tcos;
275 const fftsample_t *tsin = s->tsin;
276 const fftsample_t *in1, *in2;
277 struct fft_complex *z = (struct fft_complex *)output;
286 in2 = input + n2 - 1;
287 for (k = 0; k < n4; k++) {
289 CMUL(z[j].re, z[j].im, *in2, *in1, tcos[k], tsin[k]);
295 /* post rotation + reordering */
296 for (k = 0; k < n8; k++) {
297 fftsample_t r0, i0, r1, i1;
298 CMUL(r0, i1, z[n8 - k - 1].im, z[n8 - k - 1].re,
299 tsin[n8 - k - 1], tcos[n8 - k - 1]);
300 CMUL(r1, i0, z[n8 + k].im, z[n8 + k].re, tsin[n8 + k],
302 z[n8 - k - 1].re = r0;
303 z[n8 - k - 1].im = i0;
310 * Compute the inverse MDCT.
312 * \param ctx The initialized context structure.
313 * \param output N samples.
314 * \param input N/2 samples.
316 * \sa \ref imdct_init().
318 void imdct(struct mdct_context *ctx, float *output, const float *input)
321 int n = 1 << ctx->nbits;
325 imdct_half(ctx, output + n4, input);
327 for (k = 0; k < n4; k++) {
328 output[k] = -output[n2 - k - 1];
329 output[n - k - 1] = output[n2 + k];
333 static int fft_init(struct fft_context *s, int nbits)
337 if (nbits < 2 || nbits > 16)
338 return -E_FFT_BAD_PARAMS;
342 s->revtab = para_malloc(n * sizeof(uint16_t));
343 for (j = 4; j <= nbits; j++) {
345 double freq = 2 * M_PI / k;
346 fftsample_t *tab = cos_tabs[j - 4];
347 for (i = 0; i <= k / 4; i++)
348 tab[i] = cos(i * freq);
349 for (i = 1; i < k / 4; i++)
350 tab[k / 2 - i] = tab[i];
352 for (i = 0; i < n; i++)
353 s->revtab[-split_radix_permutation(i, n) & (n - 1)] = i;
358 * Initialize the inverse modified cosine transform.
360 * \param nbits The number of bits to use (4 <= \a nbits <= 18).
362 * \param result Opaque structure that must be passed to \ref imdct().
366 int imdct_init(int nbits, struct mdct_context **result)
370 struct mdct_context *s;
372 s = para_calloc(sizeof(*s));
377 s->tcos = para_malloc(n4 * sizeof(fftsample_t));
378 s->tsin = para_malloc(n4 * sizeof(fftsample_t));
380 for (i = 0; i < n4; i++) {
381 alpha = 2 * M_PI * (i + 1.0 / 8.0) / n;
382 s->tcos[i] = -cos(alpha);
383 s->tsin[i] = -sin(alpha);
385 ret = fft_init(&s->fft, s->nbits - 2);
398 * Deallocate imdct resources.
400 * \param ctx The pointer obtained by imdct_init().
402 void imdct_end(struct mdct_context *ctx)
406 free(ctx->fft.revtab);