* \file imdct.c Inverse modified discrete cosine transform.
*/
-#include <inttypes.h>
#include <math.h>
-#include <string.h>
-#include <stdlib.h>
#include <regex.h>
#include "para.h"
typedef float fftsample_t;
+/** Canonical representation of a complex number. */
struct fft_complex {
- fftsample_t re, im;
+ /** Real part. */
+ fftsample_t re;
+ /** Imaginary part. */
+ fftsample_t im;
};
+/** FFT Lookup table. */
struct fft_context {
+ /** Number of bits of this instance of the FFT. */
int nbits;
+ /** The lookup table for cosine values. */
uint16_t *revtab;
};
struct mdct_context {
- /** Size of MDCT (i.e. number of input data * 2). */
+ /** Size of MDCT (number of input data * 2). */
int n;
/** n = 2^n bits. */
int nbits;
- /** pre/post rotation tables */
+ /** Cosine table for pre/post rotation. */
fftsample_t *tcos;
+ /** Sine table for pre/post rotation. */
fftsample_t *tsin;
+ /** The context for the underlying fast Fourier transform. */
struct fft_context fft;
};
-/* cos(2*pi*x/n) for 0<=x<=n/4, followed by its reverse */
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_16[8]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_32[16]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_64[32]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_128[64]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_256[128]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_512[256]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_1024[512]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_2048[1024]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_4096[2048]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_8192[4096]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_16384[8192]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_32768[16384]);
-DECLARE_ALIGNED_16(fftsample_t, ff_cos_65536[32768]);
-
-static fftsample_t *ff_cos_tabs[] = {
- ff_cos_16, ff_cos_32, ff_cos_64, ff_cos_128, ff_cos_256,
- ff_cos_512, ff_cos_1024, ff_cos_2048, ff_cos_4096, ff_cos_8192,
- ff_cos_16384, ff_cos_32768, ff_cos_65536,
+/** \cond cosine_tabs */
+
+/* cos(2 * pi * x / n) for 0 <= x <= n / 4, followed by its reverse */
+#define COSINE_TAB(n) static fftsample_t cos_ ## n[n / 2] __a_aligned(16)
+
+COSINE_TAB(16);
+COSINE_TAB(32);
+COSINE_TAB(64);
+COSINE_TAB(128);
+COSINE_TAB(256);
+COSINE_TAB(512);
+COSINE_TAB(1024);
+COSINE_TAB(2048);
+COSINE_TAB(4096);
+COSINE_TAB(8192);
+COSINE_TAB(16384);
+COSINE_TAB(32768);
+COSINE_TAB(65536);
+
+static fftsample_t *cos_tabs[] = {
+ cos_16, cos_32, cos_64, cos_128, cos_256, cos_512, cos_1024, cos_2048,
+ cos_4096, cos_8192, cos_16384, cos_32768, cos_65536,
};
+/** \endcond cosine_tabs */
-static int split_radix_permutation(int i, int n)
+__a_const static int split_radix_permutation(int i, int n)
{
int m;
if (n <= 2)
return split_radix_permutation(i, m) * 4 - 1;
}
-#define SQRTHALF (float)0.70710678118654752440 /* 1/sqrt(2) */
-
-#define BF(x,y,a,b) {\
- x = a - b;\
- y = a + b;\
+#define BF(x, y, a, b) {\
+ x = a - b;\
+ y = a + b;\
}
-#define BUTTERFLIES(a0,a1,a2,a3) {\
- BF(t3, t5, t5, t1);\
- BF(a2.re, a0.re, a0.re, t5);\
- BF(a3.im, a1.im, a1.im, t3);\
- BF(t4, t6, t2, t6);\
- BF(a3.re, a1.re, a1.re, t4);\
- BF(a2.im, a0.im, a0.im, t6);\
+#define BUTTERFLIES(a0, a1, a2, a3) {\
+ BF(t3, t5, t5, t1);\
+ BF(a2.re, a0.re, a0.re, t5);\
+ BF(a3.im, a1.im, a1.im, t3);\
+ BF(t4, t6, t2, t6);\
+ BF(a3.re, a1.re, a1.re, t4);\
+ BF(a2.im, a0.im, a0.im, t6);\
}
-// force loading all the inputs before storing any.
-// this is slightly slower for small data, but avoids store->load aliasing
-// for addresses separated by large powers of 2.
-#define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
- fftsample_t r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
- BF(t3, t5, t5, t1);\
- BF(a2.re, a0.re, r0, t5);\
- BF(a3.im, a1.im, i1, t3);\
- BF(t4, t6, t2, t6);\
- BF(a3.re, a1.re, r1, t4);\
- BF(a2.im, a0.im, i0, t6);\
+/*
+ * Force loading all the inputs before storing any. This is slightly slower for
+ * small data, but avoids store->load aliasing for addresses separated by large
+ * powers of 2.
+ */
+#define BUTTERFLIES_BIG(a0, a1, a2, a3) {\
+ fftsample_t r0 = a0.re, i0 = a0.im, r1 = a1.re, i1 = a1.im;\
+ BF(t3, t5, t5, t1);\
+ BF(a2.re, a0.re, r0, t5);\
+ BF(a3.im, a1.im, i1, t3);\
+ BF(t4, t6, t2, t6);\
+ BF(a3.re, a1.re, r1, t4);\
+ BF(a2.im, a0.im, i0, t6);\
}
-#define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
- t1 = a2.re * wre + a2.im * wim;\
- t2 = a2.im * wre - a2.re * wim;\
- t5 = a3.re * wre - a3.im * wim;\
- t6 = a3.im * wre + a3.re * wim;\
- BUTTERFLIES(a0,a1,a2,a3)\
+#define TRANSFORM(a0, a1, a2, a3, wre,wim) {\
+ t1 = a2.re * wre + a2.im * wim;\
+ t2 = a2.im * wre - a2.re * wim;\
+ t5 = a3.re * wre - a3.im * wim;\
+ t6 = a3.im * wre + a3.re * wim;\
+ BUTTERFLIES(a0, a1, a2, a3)\
}
-#define TRANSFORM_ZERO(a0,a1,a2,a3) {\
- t1 = a2.re;\
- t2 = a2.im;\
- t5 = a3.re;\
- t6 = a3.im;\
- BUTTERFLIES(a0,a1,a2,a3)\
+#define TRANSFORM_ZERO(a0, a1, a2, a3) {\
+ t1 = a2.re;\
+ t2 = a2.im;\
+ t5 = a3.re;\
+ t6 = a3.im;\
+ BUTTERFLIES(a0, a1, a2, a3)\
}
-/* z[0...8n-1], w[1...2n-1] */
+/* z[0...8n - 1], w[1...2n - 1] */
#define PASS(name)\
static void name(struct fft_complex *z, const fftsample_t *wre, unsigned int n)\
{\
- fftsample_t t1, t2, t3, t4, t5, t6;\
- int o1 = 2*n;\
- int o2 = 4*n;\
- int o3 = 6*n;\
- const fftsample_t *wim = wre+o1;\
- n--;\
+ fftsample_t t1, t2, t3, t4, t5, t6;\
+ int o1 = 2 * n;\
+ int o2 = 4 * n;\
+ int o3 = 6 * n;\
+ const fftsample_t *wim = wre + o1;\
+ n--;\
\
- TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
- TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
- do {\
- z += 2;\
- wre += 2;\
- wim -= 2;\
- TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
- TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
- } while(--n);\
+ TRANSFORM_ZERO(z[0], z[o1], z[o2], z[o3]);\
+ TRANSFORM(z[1], z[o1 + 1], z[o2 + 1], z[o3 + 1], wre[1], wim[-1]);\
+ do {\
+ z += 2;\
+ wre += 2;\
+ wim -= 2;\
+ TRANSFORM(z[0], z[o1], z[o2], z[o3], wre[0], wim[0]);\
+ TRANSFORM(z[1], z[o1 + 1], z[o2 + 1], z[o3 + 1], wre[1], wim[-1]);\
+ } while (--n);\
}
PASS(pass)
#undef BUTTERFLIES
#define BUTTERFLIES BUTTERFLIES_BIG
-#define DECL_FFT(n,n2,n4)\
+#define DECL_FFT(n, n2, n4)\
static void fft##n(struct fft_complex *z)\
{\
- fft##n2(z);\
- fft##n4(z+n4*2);\
- fft##n4(z+n4*3);\
- pass(z,ff_cos_##n,n4/2);\
+ fft ## n2(z);\
+ fft ## n4(z + n4 * 2);\
+ fft ## n4(z + n4 * 3);\
+ pass(z, cos_ ## n, n4 / 2);\
}
+
static void fft4(struct fft_complex *z)
{
fftsample_t t1, t2, t3, t4, t5, t6, t7, t8;
BF(z[6].re, z[2].re, z[2].re, t7);
BF(z[6].im, z[2].im, z[2].im, t8);
- TRANSFORM(z[1], z[3], z[5], z[7], SQRTHALF, SQRTHALF);
+ TRANSFORM(z[1], z[3], z[5], z[7], M_SQRT1_2, M_SQRT1_2);
}
static void fft16(struct fft_complex *z)
fft4(z + 12);
TRANSFORM_ZERO(z[0], z[4], z[8], z[12]);
- TRANSFORM(z[2], z[6], z[10], z[14], SQRTHALF, SQRTHALF);
- TRANSFORM(z[1], z[5], z[9], z[13], ff_cos_16[1], ff_cos_16[3]);
- TRANSFORM(z[3], z[7], z[11], z[15], ff_cos_16[3], ff_cos_16[1]);
+ TRANSFORM(z[2], z[6], z[10], z[14], M_SQRT1_2, M_SQRT1_2);
+ TRANSFORM(z[1], z[5], z[9], z[13], cos_16[1], cos_16[3]);
+ TRANSFORM(z[3], z[7], z[11], z[15], cos_16[3], cos_16[1]);
}
DECL_FFT(32, 16, 8)
fft(&s->fft, z);
/* post rotation + reordering */
- output += n4;
for (k = 0; k < n8; k++) {
fftsample_t r0, i0, r1, i1;
CMUL(r0, i1, z[n8 - k - 1].im, z[n8 - k - 1].re,
for (j = 4; j <= nbits; j++) {
int k = 1 << j;
double freq = 2 * M_PI / k;
- fftsample_t *tab = ff_cos_tabs[j - 4];
+ fftsample_t *tab = cos_tabs[j - 4];
for (i = 0; i <= k / 4; i++)
tab[i] = cos(i * freq);
for (i = 1; i < k / 4; i++)
return ret;
}
+/**
+ * Deallocate imdct resources.
+ *
+ * \param ctx The pointer obtained by imdct_init().
+ */
void imdct_end(struct mdct_context *ctx)
{
free(ctx->tcos);