ref: 2e26b82ec24060f61bc553d1f36efad49d97167e
parent: 8e02afcfff4d1e4ddbf06a5b6878d49d6c3fccb5
author: Jean-Marc Valin <[email protected]>
date: Fri Dec 20 18:13:29 EST 2013
Moves the bitrev step to the IMDCT pre-rotation
--- a/celt/kiss_fft.c
+++ b/celt/kiss_fft.c
@@ -665,7 +665,7 @@
}
}
-void opus_ifft(const kiss_fft_state *st,const kiss_fft_cpx *fin,kiss_fft_cpx *fout)
+void opus_ifft_impl(const kiss_fft_state *st,kiss_fft_cpx *fout)
{
int m2, m;
int p;
@@ -676,11 +676,6 @@
/* st->shift can be -1 */
shift = st->shift>0 ? st->shift : 0;
- celt_assert2 (fin != fout, "In-place FFT not supported");
- /* Bit-reverse the input */
- for (i=0;i<st->nfft;i++)
- fout[st->bitrev[i]] = fin[i];
-
fstride[0] = 1;
L=0;
do {
@@ -717,3 +712,14 @@
}
}
+#ifdef TEST_UNIT_DFT_C
+void opus_ifft(const kiss_fft_state *st,const kiss_fft_cpx *fin,kiss_fft_cpx *fout)
+{
+ int i;
+ celt_assert2 (fin != fout, "In-place FFT not supported");
+ /* Bit-reverse the input */
+ for (i=0;i<st->nfft;i++)
+ fout[st->bitrev[i]] = fin[i];
+ opus_ifft_impl(st, fout);
+}
+#endif
--- a/celt/kiss_fft.h
+++ b/celt/kiss_fft.h
@@ -130,6 +130,8 @@
void opus_fft(const kiss_fft_state *cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout);
void opus_ifft(const kiss_fft_state *cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout);
+void opus_ifft_impl(const kiss_fft_state *st,kiss_fft_cpx *fout);
+
void opus_fft_free(const kiss_fft_state *cfg);
#ifdef __cplusplus
--- a/celt/mdct.c
+++ b/celt/mdct.c
@@ -214,13 +214,13 @@
int i;
int N, N2, N4;
kiss_twiddle_scalar sine;
- VARDECL(kiss_fft_scalar, f2);
+ VARDECL(kiss_fft_cpx, f2);
SAVE_STACK;
N = l->n;
N >>= shift;
N2 = N>>1;
N4 = N>>2;
- ALLOC(f2, N2, kiss_fft_scalar);
+ ALLOC(f2, N4, kiss_fft_cpx);
/* sin(x) ~= x here */
#ifdef FIXED_POINT
sine = TRIG_UPSCALE*(QCONST16(0.7853981f, 15)+N2)/N;
@@ -233,23 +233,26 @@
/* Temp pointers to make it really clear to the compiler what we're doing */
const kiss_fft_scalar * OPUS_RESTRICT xp1 = in;
const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+stride*(N2-1);
- kiss_fft_scalar * OPUS_RESTRICT yp = f2;
- const kiss_twiddle_scalar *t = &l->trig[0];
+ kiss_fft_cpx * OPUS_RESTRICT yp = f2;
+ const kiss_twiddle_scalar * OPUS_RESTRICT t = &l->trig[0];
+ const opus_int16 * OPUS_RESTRICT bitrev = l->kfft[shift]->bitrev;
for(i=0;i<N4;i++)
{
kiss_fft_scalar yr, yi;
+ kiss_fft_cpx yc;
yr = -S_MUL(*xp2, t[i<<shift]) + S_MUL(*xp1,t[(N4-i)<<shift]);
yi = -S_MUL(*xp2, t[(N4-i)<<shift]) - S_MUL(*xp1,t[i<<shift]);
/* works because the cos is nearly one */
- *yp++ = yr - S_MUL(yi,sine);
- *yp++ = yi + S_MUL(yr,sine);
+ yc.r = yr - S_MUL(yi,sine);
+ yc.i = yi + S_MUL(yr,sine);
+ /* Storing the pre-rotation directly in the bitrev order. */
+ yp[*bitrev++] = yc;
xp1+=2*stride;
xp2-=2*stride;
}
}
- /* Inverse N/4 complex FFT. This one should *not* downscale even in fixed-point */
- opus_ifft(l->kfft[shift], (kiss_fft_cpx *)f2, (kiss_fft_cpx *)(out+(overlap>>1)));
+ opus_ifft_impl(l->kfft[shift], f2);
/* Post-rotate and de-shuffle from both ends of the buffer at once to make
it in-place. */
@@ -263,15 +266,15 @@
{
kiss_fft_scalar re, im, yr, yi;
kiss_twiddle_scalar t0, t1;
- re = yp0[0];
- im = yp0[1];
+ re = f2[i].r;
+ im = f2[i].i;
t0 = t[i<<shift];
t1 = t[(N4-i)<<shift];
/* We'd scale up by 2 here, but instead it's done when mixing the windows */
yr = S_MUL(re,t0) - S_MUL(im,t1);
yi = S_MUL(im,t0) + S_MUL(re,t1);
- re = yp1[0];
- im = yp1[1];
+ re = f2[N4-i-1].r;
+ im = f2[N4-i-1].i;
/* works because the cos is nearly one */
yp0[0] = -(yr - S_MUL(yi,sine));
yp1[1] = yi + S_MUL(yr,sine);
--- a/celt/tests/test_unit_dft.c
+++ b/celt/tests/test_unit_dft.c
@@ -38,6 +38,7 @@
#include <stdio.h>
#define CELT_C
+#define TEST_UNIT_DFT_C
#include "stack_alloc.h"
#include "kiss_fft.h"
#include "kiss_fft.c"