devs/foxbox
1
0
Fork 0
forked from len0rd/rockbox
Code Pull requests Activity

reuse the largest sin and cos tables for smaller sizes

Browse source 
git-svn-id: svn://svn.rockbox.org/rockbox/trunk@14786 a1c6a512-1295-4272-9138-f99709370657
This commit is contained in:
Marcoen Hirschberg 2007-09-20 17:35:36 +00:00
parent 46c2363889
commit 6475ad7821
1 changed files with 24 additions and 32 deletions
Download patch file Download diff file Expand all files Collapse all files

56
apps/codecs/libwma/mdct.c
View file

@ -22,10 +22,7 @@
#include "wmafixed.h" #include "wmafixed.h"
#include "fft.h" #include "fft.h"
fixed32 *tcosarray[5], *tsinarray[5]; fixed32 tcos0[1024], tsin0[1024]; //these are the sin and cos rotations used by the MDCT
fixed32 tcos0[1024], tcos1[512], tcos2[256], tcos3[128], tcos4[64]; //these are the sin and cos rotations used by the MDCT
fixed32 tsin0[1024], tsin1[512], tsin2[256], tsin3[128], tsin4[64];
uint16_t revtab0[1024]; uint16_t revtab0[1024];
/** /**
@ -33,32 +30,16 @@ uint16_t revtab0[1024];
*/ */
int ff_mdct_init(MDCTContext *s, int nbits, int inverse) int ff_mdct_init(MDCTContext *s, int nbits, int inverse)
{ {
int n, n4, i; int n;
// fixed32 alpha; // fixed32 alpha;
memset(s, 0, sizeof(*s)); memset(s, 0, sizeof(*s));
n = 1 << nbits; //nbits ranges from 12 to 8 inclusive n = 1 << nbits; //nbits ranges from 12 to 8 inclusive
s->nbits = nbits; s->nbits = nbits;
s->n = n; s->n = n;
n4 = n >> 2;
s->tcos = tcosarray[12-nbits];
s->tsin = tsinarray[12-nbits];
for(i=0;i<n4;i++)
{
//fixed32 pi2 = fixmul32(0x20000, M_PI_F);
fixed32 ip = itofix32(i) + 0x2000;
ip = ip >> nbits;
//ip = fixdiv32(ip,itofix32(n)); // PJJ optimize
//alpha = fixmul32(TWO_M_PI_F, ip);
//s->tcos[i] = -fixcos32(alpha); //alpha between 0 and pi/2
//s->tsin[i] = -fixsin32(alpha);
s->tsin[i] = - fsincos(ip<<16, &(s->tcos[i])); //I can't remember why this works, but it seems to agree for ~24 bits, maybe more!
s->tcos[i] *=-1;
}
(&s->fft)->nbits = nbits-2; (&s->fft)->nbits = nbits-2;
(&s->fft)->inverse = inverse; (&s->fft)->inverse = inverse;
return 0; return 0;
@ -76,8 +57,6 @@ void ff_imdct_calc(MDCTContext *s,
fixed32 *input) fixed32 *input)
{ {
int k, n8, n4, n2, n, j,scale; int k, n8, n4, n2, n, j,scale;
const fixed32 *tcos = s->tcos;
const fixed32 *tsin = s->tsin;
const fixed32 *in1, *in2; const fixed32 *in1, *in2;
FFTComplex *z1 = (FFTComplex *)output; FFTComplex *z1 = (FFTComplex *)output;
FFTComplex *z2 = (FFTComplex *)input; FFTComplex *z2 = (FFTComplex *)input;
@ -96,19 +75,21 @@ void ff_imdct_calc(MDCTContext *s,
for(k = 0; k < n4; k++) for(k = 0; k < n4; k++)
{ {
j=revtab0[k<<revtabshift]; int kshift = k<<revtabshift;
CMUL(&z1[j].re, &z1[j].im, *in2, *in1, tcos[k], tsin[k]); j=revtab0[kshift];
CMUL(&z1[j].re, &z1[j].im, *in2, *in1, tcos0[kshift], tsin0[kshift]);
in1 += 2; in1 += 2;
in2 -= 2; in2 -= 2;
} }
scale = fft_calc_unscaled(&s->fft, z1); scale = fft_calc_unscaled(&s->fft, z1);
/* post rotation + reordering */ /* post rotation + reordering */
for(k = 0; k < n4; k++) for(k = 0; k < n4; k++)
{ {
CMUL(&z2[k].re, &z2[k].im, (z1[k].re), (z1[k].im), tcos[k], tsin[k]); int kshift = k<<revtabshift;
CMUL(&z2[k].re, &z2[k].im, (z1[k].re), (z1[k].im), tcos0[kshift], tsin0[kshift]);
} }
for(k = 0; k < n8; k++) for(k = 0; k < n8; k++)
@ -140,10 +121,6 @@ void ff_imdct_calc(MDCTContext *s,
int mdct_init_global(void) int mdct_init_global(void)
{ {
int i,j,m; int i,j,m;
/* init MDCT */
/*TODO: figure out how to fold this up into one array*/
tcosarray[0] = tcos0; tcosarray[1] = tcos1; tcosarray[2] = tcos2; tcosarray[3] = tcos3;tcosarray[4] = tcos4;
tsinarray[0] = tsin0; tsinarray[1] = tsin1; tsinarray[2] = tsin2; tsinarray[3] = tsin3;tsinarray[4] = tsin4;
/* init the MDCT bit reverse table here rather then in fft_init */ /* init the MDCT bit reverse table here rather then in fft_init */
for(i=0;i<1024;i++) /*hard coded to a 2048 bit rotation*/ for(i=0;i<1024;i++) /*hard coded to a 2048 bit rotation*/
@ -157,6 +134,21 @@ int mdct_init_global(void)
revtab0[i]=m; revtab0[i]=m;
} }
for(i=0;i<1024;i++)
{
//fixed32 pi2 = fixmul32(0x20000, M_PI_F);
fixed32 ip = itofix32(i) + 0x2000;
ip = ip >> 12;
//ip = fixdiv32(ip,itofix32(n)); // PJJ optimize
//alpha = fixmul32(TWO_M_PI_F, ip);
//s->tcos[i] = -fixcos32(alpha); //alpha between 0 and pi/2
//s->tsin[i] = -fixsin32(alpha);
//I can't remember why this works, but it seems to agree for ~24 bits, maybe more!
tsin0[i] = - fsincos(ip<<16, &(tcos0[i]));
tcos0[i] *=-1;
}
fft_init_global(); fft_init_global();
return 0; return 0;