void a52_imdct_512 (sample_t * data, sample_t * delay, sample_t bias)
{
int i, k;
sample_t t_r, t_i, a_r, a_i, b_r, b_i, w_1, w_2;
const sample_t * window = a52_imdct_window;
complex_t buf[128];
for (i = 0; i < 128; i++) {
k = fftorder[i];
t_r = pre1[i].real;
t_i = pre1[i].imag;
BUTTERFLY_0 (buf[i].real, buf[i].imag, t_r, t_i, data[k], data[255-k]);
}
ifft128 (buf);
/* Post IFFT complex multiply plus IFFT complex conjugate*/
/* Window and convert to real valued signal */
for (i = 0; i < 64; i++) {
/* y[n] = z[n] * (xcos1[n] + j * xsin1[n]) ; */
t_r = post1[i].real;
t_i = post1[i].imag;
BUTTERFLY_0 (a_r, a_i, t_i, t_r, buf[i].imag, buf[i].real);
BUTTERFLY_0 (b_r, b_i, t_r, t_i, buf[127-i].imag, buf[127-i].real);
w_1 = window[2*i];
w_2 = window[255-2*i];
BUTTERFLY_B (data[255-2*i], data[2*i], w_2, w_1, a_r, delay[2*i]);
delay[2*i] = a_i;
w_1 = window[2*i+1];
w_2 = window[254-2*i];
BUTTERFLY_B (data[2*i+1], data[254-2*i], w_1, w_2, b_r, delay[2*i+1]);
delay[2*i+1] = b_i;
}
}
void a52_imdct_256 (sample_t * data, sample_t * delay, sample_t bias)
{
int i, k;
sample_t t_r, t_i, a_r, a_i, b_r, b_i, c_r, c_i, d_r, d_i, w_1, w_2;
const sample_t * window = a52_imdct_window;
complex_t buf1[64], buf2[64];
/* Pre IFFT complex multiply plus IFFT cmplx conjugate */
for (i = 0; i < 64; i++) {
k = fftorder[i];
t_r = pre2[i].real;
t_i = pre2[i].imag;
BUTTERFLY_0 (buf1[i].real, buf1[i].imag, t_r, t_i, data[k], data[254-k]);
BUTTERFLY_0 (buf2[i].real, buf2[i].imag, t_r, t_i, data[k+1], data[255-k]);
}
ifft64 (buf1);
ifft64 (buf2);
/* Post IFFT complex multiply */
/* Window and convert to real valued signal */
for (i = 0; i < 32; i++) {
/* y1[n] = z1[n] * (xcos2[n] + j * xs in2[n]) ; */
t_r = post2[i].real;
t_i = post2[i].imag;
BUTTERFLY_0 (a_r, a_i, t_i, t_r, buf1[i].imag, buf1[i].real);
BUTTERFLY_0 (b_r, b_i, t_r, t_i, buf1[63-i].imag, buf1[63-i].real);
BUTTERFLY_0 (c_r, c_i, t_i, t_r, buf2[i].imag, buf2[i].real);
BUTTERFLY_0 (d_r, d_i, t_r, t_i, buf2[63-i].imag, buf2[63-i].real);
w_1 = window[2*i];
w_2 = window[255-2*i];
BUTTERFLY_B (data[255-2*i], data[2*i], w_2, w_1, a_r, delay[2*i]);
delay[2*i] = c_i;
w_1 = window[128+2*i];
w_2 = window[127-2*i];
BUTTERFLY_B (data[128+2*i], data[127-2*i], w_1, w_2, a_i, delay[127-2*i]);
delay[127-2*i] = c_r;
w_1 = window[2*i+1];
w_2 = window[254-2*i];
BUTTERFLY_B (data[254-2*i], data[2*i+1], w_2, w_1, b_i, delay[2*i+1]);
delay[2*i+1] = d_r;
w_1 = window[129+2*i];
w_2 = window[126-2*i];
BUTTERFLY_B (data[129+2*i], data[126-2*i], w_1, w_2, b_r, delay[126-2*i]);
delay[126-2*i] = d_i;
}
}
void fft64_2way( void *a )
{
__m256i* const A = a;
register __m256i X0, X1, X2, X3, X4, X5, X6, X7;
#define X(i) X##i
X0 = A[0];
X1 = A[1];
X2 = A[2];
X3 = A[3];
X4 = A[4];
X5 = A[5];
X6 = A[6];
X7 = A[7];
#define DO_REDUCE(i) X(i) = REDUCE( X(i) )
// Begin with 8 parallels DIF FFT_8
//
// FFT_8 using w=4 as 8th root of unity
// Unrolled decimation in frequency (DIF) radix-2 NTT.
// Output data is in revbin_permuted order.
static const int w[] = {0, 2, 4, 6};
// __m256i *Twiddle = (__m256i*)FFT64_Twiddle;
#define BUTTERFLY_0( i,j ) \
do { \
__m256i v = X(j); \
X(j) = _mm256_add_epi16( X(i), X(j) ); \
X(i) = _mm256_sub_epi16( X(i), v ); \
} while(0)
#define BUTTERFLY_N( i,j,n ) \
do { \
__m256i v = X(j); \
X(j) = _mm256_add_epi16( X(i), X(j) ); \
X(i) = _mm256_slli_epi16( _mm256_sub_epi16( X(i), v ), w[n] ); \
} while(0)
BUTTERFLY_0( 0, 4 );
BUTTERFLY_N( 1, 5, 1 );
BUTTERFLY_N( 2, 6, 2 );
BUTTERFLY_N( 3, 7, 3 );
DO_REDUCE( 2 );
DO_REDUCE( 3 );
BUTTERFLY_0( 0, 2 );
BUTTERFLY_0( 4, 6 );
BUTTERFLY_N( 1, 3, 2 );
BUTTERFLY_N( 5, 7, 2 );
DO_REDUCE( 1 );
BUTTERFLY_0( 0, 1 );
BUTTERFLY_0( 2, 3 );
BUTTERFLY_0( 4, 5 );
BUTTERFLY_0( 6, 7 );
/* We don't need to reduce X(7) */
DO_REDUCE_FULL_S( 0 );
DO_REDUCE_FULL_S( 1 );
DO_REDUCE_FULL_S( 2 );
DO_REDUCE_FULL_S( 3 );
DO_REDUCE_FULL_S( 4 );
DO_REDUCE_FULL_S( 5 );
DO_REDUCE_FULL_S( 6 );
#undef BUTTERFLY_0
#undef BUTTERFLY_N
// Multiply by twiddle factors
X(6) = _mm256_mullo_epi16( X(6), FFT64_Twiddle[0].m256i );
X(5) = _mm256_mullo_epi16( X(5), FFT64_Twiddle[1].m256i );
X(4) = _mm256_mullo_epi16( X(4), FFT64_Twiddle[2].m256i );
X(3) = _mm256_mullo_epi16( X(3), FFT64_Twiddle[3].m256i );
X(2) = _mm256_mullo_epi16( X(2), FFT64_Twiddle[4].m256i );
X(1) = _mm256_mullo_epi16( X(1), FFT64_Twiddle[5].m256i );
X(0) = _mm256_mullo_epi16( X(0), FFT64_Twiddle[6].m256i );
// Transpose the FFT state with a revbin order permutation
// on the rows and the column.
// This will make the full FFT_64 in order.
#define INTERLEAVE(i,j) \
do { \
__m256i t1= X(i); \
__m256i t2= X(j); \
X(i) = _mm256_unpacklo_epi16( t1, t2 ); \
X(j) = _mm256_unpackhi_epi16( t1, t2 ); \
} while(0)
INTERLEAVE( 1, 0 );
INTERLEAVE( 3, 2 );
INTERLEAVE( 5, 4 );
INTERLEAVE( 7, 6 );
INTERLEAVE( 2, 0 );
INTERLEAVE( 3, 1 );
INTERLEAVE( 6, 4 );
INTERLEAVE( 7, 5 );
INTERLEAVE( 4, 0 );
INTERLEAVE( 5, 1 );
INTERLEAVE( 6, 2 );
INTERLEAVE( 7, 3 );
#undef INTERLEAVE
//Finish with 8 parallels DIT FFT_8
//FFT_8 using w=4 as 8th root of unity
// Unrolled decimation in time (DIT) radix-2 NTT.
// Input data is in revbin_permuted order.
#define BUTTERFLY_0( i,j ) \
do { \
__m256i u = X(j); \
X(j) = _mm256_sub_epi16( X(j), X(i) ); \
X(i) = _mm256_add_epi16( u, X(i) ); \
} while(0)
#define BUTTERFLY_N( i,j,n ) \
do { \
__m256i u = X(j); \
X(i) = _mm256_slli_epi16( X(i), w[n] ); \
X(j) = _mm256_sub_epi16( X(j), X(i) ); \
X(i) = _mm256_add_epi16( u, X(i) ); \
} while(0)
DO_REDUCE( 0 );
DO_REDUCE( 1 );
DO_REDUCE( 2 );
DO_REDUCE( 3 );
DO_REDUCE( 4 );
DO_REDUCE( 5 );
DO_REDUCE( 6 );
DO_REDUCE( 7 );
BUTTERFLY_0( 0, 1 );
BUTTERFLY_0( 2, 3 );
BUTTERFLY_0( 4, 5 );
BUTTERFLY_0( 6, 7 );
BUTTERFLY_0( 0, 2 );
BUTTERFLY_0( 4, 6 );
BUTTERFLY_N( 1, 3, 2 );
BUTTERFLY_N( 5, 7, 2 );
DO_REDUCE( 3 );
BUTTERFLY_0( 0, 4 );
BUTTERFLY_N( 1, 5, 1 );
BUTTERFLY_N( 2, 6, 2 );
BUTTERFLY_N( 3, 7, 3 );
DO_REDUCE_FULL_S( 0 );
DO_REDUCE_FULL_S( 1 );
DO_REDUCE_FULL_S( 2 );
DO_REDUCE_FULL_S( 3 );
DO_REDUCE_FULL_S( 4 );
DO_REDUCE_FULL_S( 5 );
DO_REDUCE_FULL_S( 6 );
DO_REDUCE_FULL_S( 7 );
#undef BUTTERFLY
A[0] = X0;
A[1] = X1;
A[2] = X2;
A[3] = X3;
A[4] = X4;
A[5] = X5;
A[6] = X6;
A[7] = X7;
#undef X
}
