typedef float fftsample_t;
-#define DECLARE_ALIGNED(n,t,v) t v __attribute__ ((aligned (n)))
-#define DECLARE_ALIGNED_16(t, v) DECLARE_ALIGNED(16, t, v)
-#define M_SQRT1_2 0.70710678118654752440 /* 1/sqrt(2) */
-
+/** 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;
- int inverse;
+ /** The lookup table for cosine values. */
uint16_t *revtab;
- struct fft_complex *exptab;
- struct fft_complex *exptab1; /* only used by SSE code */
- struct fft_complex *tmp_buf;
};
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,
+/** 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,
};
-static int split_radix_permutation(int i, int n, int inverse)
+__a_const static int split_radix_permutation(int i, int n)
{
int m;
if (n <= 2)
return i & 1;
m = n >> 1;
- if (!(i & m))
- return split_radix_permutation(i, m, inverse) * 2;
+ if ((i & m) == 0)
+ return split_radix_permutation(i, m) * 2;
m >>= 1;
- if (inverse == !(i & m))
- return split_radix_permutation(i, m, inverse) * 4 + 1;
+ if ((i & m) == 0)
+ return split_radix_permutation(i, m) * 4 + 1;
else
- return split_radix_permutation(i, m, inverse) * 4 - 1;
+ return split_radix_permutation(i, m) * 4 - 1;
}
-#define sqrthalf (float)M_SQRT1_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)
}
/**
- * Compute the inverse MDCT of size N = 2^nbits.
+ * Compute the inverse MDCT.
*
+ * \param ctx The initialized context structure.
* \param output N samples.
* \param input N/2 samples.
+ *
+ * \sa \ref imdct_init().
*/
-void imdct(struct mdct_context *s, float *output, const float *input)
+void imdct(struct mdct_context *ctx, float *output, const float *input)
{
int k;
- int n = 1 << s->nbits;
+ int n = 1 << ctx->nbits;
int n2 = n >> 1;
int n4 = n >> 2;
- imdct_half(s, output + n4, input);
+ imdct_half(ctx, output + n4, input);
for (k = 0; k < n4; k++) {
output[k] = -output[n2 - k - 1];
}
}
-static int fft_init(struct fft_context *s, int nbits, int inverse)
+static int fft_init(struct fft_context *s, int nbits)
{
- int i, j, m, n;
- float alpha, c1, s1, s2;
- int split_radix = 1;
+ int i, j, n;
if (nbits < 2 || nbits > 16)
return -E_FFT_BAD_PARAMS;
s->nbits = nbits;
n = 1 << nbits;
- s->tmp_buf = NULL;
- s->exptab = para_malloc((n / 2) * sizeof(struct fft_complex));
s->revtab = para_malloc(n * sizeof(uint16_t));
- s->inverse = inverse;
-
- s2 = inverse ? 1.0 : -1.0;
-
- s->exptab1 = NULL;
-
- if (split_radix) {
- for (j = 4; j <= nbits; j++) {
- int k = 1 << j;
- double freq = 2 * M_PI / k;
- fftsample_t *tab = ff_cos_tabs[j - 4];
- for (i = 0; i <= k / 4; i++)
- tab[i] = cos(i * freq);
- for (i = 1; i < k / 4; i++)
- tab[k / 2 - i] = tab[i];
- }
- for (i = 0; i < n; i++)
- s->revtab[-split_radix_permutation(
- i, n, s->inverse) & (n - 1)] = i;
- s->tmp_buf = para_malloc(n * sizeof(struct fft_complex));
- } else {
- int np, nblocks, np2, l;
- struct fft_complex *q;
-
- for (i = 0; i < (n / 2); i++) {
- alpha = 2 * M_PI * (float) i / (float) n;
- c1 = cos(alpha);
- s1 = sin(alpha) * s2;
- s->exptab[i].re = c1;
- s->exptab[i].im = s1;
- }
-
- np = 1 << nbits;
- nblocks = np >> 3;
- np2 = np >> 1;
- s->exptab1 = para_malloc(np * 2 * sizeof(struct fft_complex));
- q = s->exptab1;
- do {
- for (l = 0; l < np2; l += 2 * nblocks) {
- *q++ = s->exptab[l];
- *q++ = s->exptab[l + nblocks];
-
- q->re = -s->exptab[l].im;
- q->im = s->exptab[l].re;
- q++;
- q->re = -s->exptab[l + nblocks].im;
- q->im = s->exptab[l + nblocks].re;
- q++;
- }
- nblocks = nblocks >> 1;
- } while (nblocks != 0);
- freep(&s->exptab);
- /* compute bit reverse table */
- for (i = 0; i < n; i++) {
- m = 0;
- for (j = 0; j < nbits; j++)
- m |= ((i >> j) & 1) << (nbits - j - 1);
- s->revtab[i] = m;
- }
+ for (j = 4; j <= nbits; j++) {
+ int k = 1 << j;
+ double freq = 2 * M_PI / k;
+ 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++)
+ tab[k / 2 - i] = tab[i];
}
- return 0;
-}
-
-static void fft_end(struct fft_context *ctx)
-{
- freep(&ctx->revtab);
- freep(&ctx->exptab);
- freep(&ctx->exptab1);
- freep(&ctx->tmp_buf);
-}
-
-DECLARE_ALIGNED(16, float, ff_sine_128[128]);
-DECLARE_ALIGNED(16, float, ff_sine_256[256]);
-DECLARE_ALIGNED(16, float, ff_sine_512[512]);
-DECLARE_ALIGNED(16, float, ff_sine_1024[1024]);
-DECLARE_ALIGNED(16, float, ff_sine_2048[2048]);
-DECLARE_ALIGNED(16, float, ff_sine_4096[4096]);
-
-float *ff_sine_windows[6] = {
- ff_sine_128, ff_sine_256, ff_sine_512, ff_sine_1024,
- ff_sine_2048, ff_sine_4096
-};
-
-// Generate a sine window.
-void sine_window_init(float *window, int n)
-{
- int i;
-
for (i = 0; i < n; i++)
- window[i] = sinf((i + 0.5) * (M_PI / (2.0 * n)));
+ s->revtab[-split_radix_permutation(i, n) & (n - 1)] = i;
+ return 0;
}
/**
- * Init MDCT or IMDCT computation.
+ * Initialize the inverse modified cosine transform.
+ *
+ * \param nbits The number of bits to use (4 <= \a nbits <= 18).
+ *
+ * \param result Opaque structure that must be passed to \ref imdct().
+ *
+ * \return Standard.
*/
-int imdct_init(int nbits, int inverse, struct mdct_context **result)
+int imdct_init(int nbits, struct mdct_context **result)
{
int ret, n, n4, i;
double alpha;
s->tcos[i] = -cos(alpha);
s->tsin[i] = -sin(alpha);
}
- ret = fft_init(&s->fft, s->nbits - 2, inverse);
+ ret = fft_init(&s->fft, s->nbits - 2);
if (ret < 0)
goto fail;
*result = s;
return ret;
}
+/**
+ * Deallocate imdct resources.
+ *
+ * \param ctx The pointer obtained by imdct_init().
+ */
void imdct_end(struct mdct_context *ctx)
{
- freep(&ctx->tcos);
- freep(&ctx->tsin);
- fft_end(&ctx->fft);
+ free(ctx->tcos);
+ free(ctx->tsin);
+ free(ctx->fft.revtab);
free(ctx);
}