diff options
| -rw-r--r-- | libcelt/Makefile.am | 9 | ||||
| -rw-r--r-- | libcelt/_kiss_fft_guts.h | 160 | ||||
| -rw-r--r-- | libcelt/kiss_fft.c | 523 | ||||
| -rw-r--r-- | libcelt/kiss_fft.h | 108 | ||||
| -rw-r--r-- | libcelt/mdct.c | 698 | ||||
| -rw-r--r-- | libcelt/mdct.h | 132 | ||||
| -rw-r--r-- | libcelt/testcelt.c | 5 |
7 files changed, 1001 insertions, 634 deletions
diff --git a/libcelt/Makefile.am b/libcelt/Makefile.am index dbaaeb0..5e0b20d 100644 --- a/libcelt/Makefile.am +++ b/libcelt/Makefile.am @@ -15,16 +15,17 @@ lib_LTLIBRARIES = libcelt.la # Sources for compilation in the library libcelt_la_SOURCES = bands.c bitrdec.c bitree.c bitrenc.c celt.c cwrs.c \ - ecintrin.h entcode.c entdec.c entenc.c fftwrap.c header.c laplace.c mdct.c modes.c \ - pitch.c psy.c quant_bands.c quant_pitch.c rangedec.c rangeenc.c rate.c \ - smallft.c vq.c + ecintrin.h entcode.c entdec.c entenc.c fftwrap.c header.c kiss_fft.c \ + laplace.c mdct.c modes.c pitch.c psy.c quant_bands.c quant_pitch.c \ + rangedec.c rangeenc.c rate.c smallft.c vq.c #noinst_HEADERS = libcelt_la_LDFLAGS = -version-info @CELT_LT_CURRENT@:@CELT_LT_REVISION@:@CELT_LT_AGE@ noinst_HEADERS = arch.h bands.h bitrdec.h bitree.h bitrenc.h cwrs.h \ - ecintrin.h entcode.h entdec.h entenc.h fftwrap.h laplace.h mdct.h \ + ecintrin.h entcode.h entdec.h entenc.h fftwrap.h kiss_fft.h \ + _kiss_fft_guts.h laplace.h mdct.h \ mfrngcod.h modes.h os_support.h pgain_table.h pitch.h psy.h \ quant_bands.h quant_pitch.h rate.h smallft.h vq.h diff --git a/libcelt/_kiss_fft_guts.h b/libcelt/_kiss_fft_guts.h new file mode 100644 index 0000000..449194b --- /dev/null +++ b/libcelt/_kiss_fft_guts.h @@ -0,0 +1,160 @@ +/* +Copyright (c) 2003-2004, Mark Borgerding + +All rights reserved. + +Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. + * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +#define MIN(a,b) ((a)<(b) ? (a):(b)) +#define MAX(a,b) ((a)>(b) ? (a):(b)) + +/* kiss_fft.h + defines kiss_fft_scalar as either short or a float type + and defines + typedef struct { kiss_fft_scalar r; kiss_fft_scalar i; }kiss_fft_cpx; */ +#include "kiss_fft.h" +//#include "math_approx.h" + +#define MAXFACTORS 32 +/* e.g. an fft of length 128 has 4 factors + as far as kissfft is concerned + 4*4*4*2 + */ + +struct kiss_fft_state{ + int nfft; + int inverse; + int factors[2*MAXFACTORS]; + kiss_fft_cpx twiddles[1]; +}; + +/* + Explanation of macros dealing with complex math: + + C_MUL(m,a,b) : m = a*b + C_FIXDIV( c , div ) : if a fixed point impl., c /= div. noop otherwise + C_SUB( res, a,b) : res = a - b + C_SUBFROM( res , a) : res -= a + C_ADDTO( res , a) : res += a + * */ +#ifdef FIXED_POINT +#include "arch.h" +# define FRACBITS 15 +# define SAMPPROD celt_int32_t +#define SAMP_MAX 32767 + +#define SAMP_MIN -SAMP_MAX + +#if defined(CHECK_OVERFLOW) +# define CHECK_OVERFLOW_OP(a,op,b) \ + if ( (SAMPPROD)(a) op (SAMPPROD)(b) > SAMP_MAX || (SAMPPROD)(a) op (SAMPPROD)(b) < SAMP_MIN ) { \ + fprintf(stderr,"WARNING:overflow @ " __FILE__ "(%d): (%d " #op" %d) = %ld\n",__LINE__,(a),(b),(SAMPPROD)(a) op (SAMPPROD)(b) ); } +#endif + + +# define smul(a,b) ( (SAMPPROD)(a)*(b) ) +# define sround( x ) (kiss_fft_scalar)( ( (x) + (1<<(FRACBITS-1)) ) >> FRACBITS ) + +# define S_MUL(a,b) sround( smul(a,b) ) + +# define C_MUL(m,a,b) \ + do{ (m).r = sround( smul((a).r,(b).r) - smul((a).i,(b).i) ); \ + (m).i = sround( smul((a).r,(b).i) + smul((a).i,(b).r) ); }while(0) + +# define C_MUL4(m,a,b) \ + do{ (m).r = PSHR32( smul((a).r,(b).r) - smul((a).i,(b).i),17 ); \ + (m).i = PSHR32( smul((a).r,(b).i) + smul((a).i,(b).r),17 ); }while(0) + +# define DIVSCALAR(x,k) \ + (x) = sround( smul( x, SAMP_MAX/k ) ) + +# define C_FIXDIV(c,div) \ + do { DIVSCALAR( (c).r , div); \ + DIVSCALAR( (c).i , div); }while (0) + +# define C_MULBYSCALAR( c, s ) \ + do{ (c).r = sround( smul( (c).r , s ) ) ;\ + (c).i = sround( smul( (c).i , s ) ) ; }while(0) + +#else /* not FIXED_POINT*/ + +# define S_MUL(a,b) ( (a)*(b) ) +#define C_MUL(m,a,b) \ + do{ (m).r = (a).r*(b).r - (a).i*(b).i;\ + (m).i = (a).r*(b).i + (a).i*(b).r; }while(0) + +#define C_MUL4(m,a,b) C_MUL(m,a,b) + +# define C_FIXDIV(c,div) /* NOOP */ +# define C_MULBYSCALAR( c, s ) \ + do{ (c).r *= (s);\ + (c).i *= (s); }while(0) +#endif + +#ifndef CHECK_OVERFLOW_OP +# define CHECK_OVERFLOW_OP(a,op,b) /* noop */ +#endif + +#define C_ADD( res, a,b)\ + do { \ + CHECK_OVERFLOW_OP((a).r,+,(b).r)\ + CHECK_OVERFLOW_OP((a).i,+,(b).i)\ + (res).r=(a).r+(b).r; (res).i=(a).i+(b).i; \ + }while(0) +#define C_SUB( res, a,b)\ + do { \ + CHECK_OVERFLOW_OP((a).r,-,(b).r)\ + CHECK_OVERFLOW_OP((a).i,-,(b).i)\ + (res).r=(a).r-(b).r; (res).i=(a).i-(b).i; \ + }while(0) +#define C_ADDTO( res , a)\ + do { \ + CHECK_OVERFLOW_OP((res).r,+,(a).r)\ + CHECK_OVERFLOW_OP((res).i,+,(a).i)\ + (res).r += (a).r; (res).i += (a).i;\ + }while(0) + +#define C_SUBFROM( res , a)\ + do {\ + CHECK_OVERFLOW_OP((res).r,-,(a).r)\ + CHECK_OVERFLOW_OP((res).i,-,(a).i)\ + (res).r -= (a).r; (res).i -= (a).i; \ + }while(0) + + +#ifdef FIXED_POINT +# define KISS_FFT_COS(phase) floor(MIN(32767,MAX(-32767,.5+32768 * cos (phase)))) +# define KISS_FFT_SIN(phase) floor(MIN(32767,MAX(-32767,.5+32768 * sin (phase)))) +# define HALF_OF(x) ((x)>>1) +#elif defined(USE_SIMD) +# define KISS_FFT_COS(phase) _mm_set1_ps( cos(phase) ) +# define KISS_FFT_SIN(phase) _mm_set1_ps( sin(phase) ) +# define HALF_OF(x) ((x)*_mm_set1_ps(.5)) +#else +# define KISS_FFT_COS(phase) (kiss_fft_scalar) cos(phase) +# define KISS_FFT_SIN(phase) (kiss_fft_scalar) sin(phase) +# define HALF_OF(x) ((x)*.5) +#endif + +#define kf_cexp(x,phase) \ + do{ \ + (x)->r = KISS_FFT_COS(phase);\ + (x)->i = KISS_FFT_SIN(phase);\ + }while(0) +#define kf_cexp2(x,phase) \ + do{ \ + (x)->r = celt_cos_norm((phase));\ + (x)->i = celt_cos_norm((phase)-32768);\ +}while(0) + + +/* a debugging function */ +#define pcpx(c)\ + fprintf(stderr,"%g + %gi\n",(double)((c)->r),(double)((c)->i) ) diff --git a/libcelt/kiss_fft.c b/libcelt/kiss_fft.c new file mode 100644 index 0000000..e254f27 --- /dev/null +++ b/libcelt/kiss_fft.c @@ -0,0 +1,523 @@ +/* +Copyright (c) 2003-2004, Mark Borgerding +Copyright (c) 2005-2007, Jean-Marc Valin + +All rights reserved. + +Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. + * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include "_kiss_fft_guts.h" +#include "arch.h" +#include "os_support.h" + +/* The guts header contains all the multiplication and addition macros that are defined for + fixed or floating point complex numbers. It also delares the kf_ internal functions. + */ + +static void kf_bfly2( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + int m, + int N, + int mm + ) +{ + kiss_fft_cpx * Fout2; + kiss_fft_cpx * tw1; + kiss_fft_cpx t; + if (!st->inverse) { + int i,j; + kiss_fft_cpx * Fout_beg = Fout; + for (i=0;i<N;i++) + { + Fout = Fout_beg + i*mm; + Fout2 = Fout + m; + tw1 = st->twiddles; + for(j=0;j<m;j++) + { + /* Almost the same as the code path below, except that we divide the input by two + (while keeping the best accuracy possible) */ + celt_word32_t tr, ti; + tr = SHR32(SUB32(MULT16_16(Fout2->r , tw1->r),MULT16_16(Fout2->i , tw1->i)), 1); + ti = SHR32(ADD32(MULT16_16(Fout2->i , tw1->r),MULT16_16(Fout2->r , tw1->i)), 1); + tw1 += fstride; + Fout2->r = PSHR32(SUB32(SHL32(EXTEND32(Fout->r), 14), tr), 15); + Fout2->i = PSHR32(SUB32(SHL32(EXTEND32(Fout->i), 14), ti), 15); + Fout->r = PSHR32(ADD32(SHL32(EXTEND32(Fout->r), 14), tr), 15); + Fout->i = PSHR32(ADD32(SHL32(EXTEND32(Fout->i), 14), ti), 15); + ++Fout2; + ++Fout; + } + } + } else { + int i,j; + kiss_fft_cpx * Fout_beg = Fout; + for (i=0;i<N;i++) + { + Fout = Fout_beg + i*mm; + Fout2 = Fout + m; + tw1 = st->twiddles; + for(j=0;j<m;j++) + { + C_MUL (t, *Fout2 , *tw1); + tw1 += fstride; + C_SUB( *Fout2 , *Fout , t ); + C_ADDTO( *Fout , t ); + ++Fout2; + ++Fout; + } + } + } +} + +static void kf_bfly4( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + int m, + int N, + int mm + ) +{ + kiss_fft_cpx *tw1,*tw2,*tw3; + kiss_fft_cpx scratch[6]; + const size_t m2=2*m; + const size_t m3=3*m; + int i, j; + + if (st->inverse) + { + kiss_fft_cpx * Fout_beg = Fout; + for (i=0;i<N;i++) + { + Fout = Fout_beg + i*mm; + tw3 = tw2 = tw1 = st->twiddles; + for (j=0;j<m;j++) + { + C_MUL(scratch[0],Fout[m] , *tw1 ); + C_MUL(scratch[1],Fout[m2] , *tw2 ); + C_MUL(scratch[2],Fout[m3] , *tw3 ); + + C_SUB( scratch[5] , *Fout, scratch[1] ); + C_ADDTO(*Fout, scratch[1]); + C_ADD( scratch[3] , scratch[0] , scratch[2] ); + C_SUB( scratch[4] , scratch[0] , scratch[2] ); + C_SUB( Fout[m2], *Fout, scratch[3] ); + tw1 += fstride; + tw2 += fstride*2; + tw3 += fstride*3; + C_ADDTO( *Fout , scratch[3] ); + + Fout[m].r = scratch[5].r - scratch[4].i; + Fout[m].i = scratch[5].i + scratch[4].r; + Fout[m3].r = scratch[5].r + scratch[4].i; + Fout[m3].i = scratch[5].i - scratch[4].r; + ++Fout; + } + } + } else + { + kiss_fft_cpx * Fout_beg = Fout; + for (i=0;i<N;i++) + { + Fout = Fout_beg + i*mm; + tw3 = tw2 = tw1 = st->twiddles; + for (j=0;j<m;j++) + { + C_MUL4(scratch[0],Fout[m] , *tw1 ); + C_MUL4(scratch[1],Fout[m2] , *tw2 ); + C_MUL4(scratch[2],Fout[m3] , *tw3 ); + + Fout->r = PSHR16(Fout->r, 2); + Fout->i = PSHR16(Fout->i, 2); + C_SUB( scratch[5] , *Fout, scratch[1] ); + C_ADDTO(*Fout, scratch[1]); + C_ADD( scratch[3] , scratch[0] , scratch[2] ); + C_SUB( scratch[4] , scratch[0] , scratch[2] ); + Fout[m2].r = PSHR16(Fout[m2].r, 2); + Fout[m2].i = PSHR16(Fout[m2].i, 2); + C_SUB( Fout[m2], *Fout, scratch[3] ); + tw1 += fstride; + tw2 += fstride*2; + tw3 += fstride*3; + C_ADDTO( *Fout , scratch[3] ); + + Fout[m].r = scratch[5].r + scratch[4].i; + Fout[m].i = scratch[5].i - scratch[4].r; + Fout[m3].r = scratch[5].r - scratch[4].i; + Fout[m3].i = scratch[5].i + scratch[4].r; + ++Fout; + } + } + } +} + +static void kf_bfly3( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + size_t m + ) +{ + size_t k=m; + const size_t m2 = 2*m; + kiss_fft_cpx *tw1,*tw2; + kiss_fft_cpx scratch[5]; + kiss_fft_cpx epi3; + epi3 = st->twiddles[fstride*m]; + + tw1=tw2=st->twiddles; + + do{ + if (!st->inverse) { + C_FIXDIV(*Fout,3); C_FIXDIV(Fout[m],3); C_FIXDIV(Fout[m2],3); + } + + C_MUL(scratch[1],Fout[m] , *tw1); + C_MUL(scratch[2],Fout[m2] , *tw2); + + C_ADD(scratch[3],scratch[1],scratch[2]); + C_SUB(scratch[0],scratch[1],scratch[2]); + tw1 += fstride; + tw2 += fstride*2; + + Fout[m].r = Fout->r - HALF_OF(scratch[3].r); + Fout[m].i = Fout->i - HALF_OF(scratch[3].i); + + C_MULBYSCALAR( scratch[0] , epi3.i ); + + C_ADDTO(*Fout,scratch[3]); + + Fout[m2].r = Fout[m].r + scratch[0].i; + Fout[m2].i = Fout[m].i - scratch[0].r; + + Fout[m].r -= scratch[0].i; + Fout[m].i += scratch[0].r; + + ++Fout; + }while(--k); +} + +static void kf_bfly5( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + int m + ) +{ + kiss_fft_cpx *Fout0,*Fout1,*Fout2,*Fout3,*Fout4; + int u; + kiss_fft_cpx scratch[13]; + kiss_fft_cpx * twiddles = st->twiddles; + kiss_fft_cpx *tw; + kiss_fft_cpx ya,yb; + ya = twiddles[fstride*m]; + yb = twiddles[fstride*2*m]; + + Fout0=Fout; + Fout1=Fout0+m; + Fout2=Fout0+2*m; + Fout3=Fout0+3*m; + Fout4=Fout0+4*m; + + tw=st->twiddles; + for ( u=0; u<m; ++u ) { + if (!st->inverse) { + C_FIXDIV( *Fout0,5); C_FIXDIV( *Fout1,5); C_FIXDIV( *Fout2,5); C_FIXDIV( *Fout3,5); C_FIXDIV( *Fout4,5); + } + scratch[0] = *Fout0; + + C_MUL(scratch[1] ,*Fout1, tw[u*fstride]); + C_MUL(scratch[2] ,*Fout2, tw[2*u*fstride]); + C_MUL(scratch[3] ,*Fout3, tw[3*u*fstride]); + C_MUL(scratch[4] ,*Fout4, tw[4*u*fstride]); + + C_ADD( scratch[7],scratch[1],scratch[4]); + C_SUB( scratch[10],scratch[1],scratch[4]); + C_ADD( scratch[8],scratch[2],scratch[3]); + C_SUB( scratch[9],scratch[2],scratch[3]); + + Fout0->r += scratch[7].r + scratch[8].r; + Fout0->i += scratch[7].i + scratch[8].i; + + scratch[5].r = scratch[0].r + S_MUL(scratch[7].r,ya.r) + S_MUL(scratch[8].r,yb.r); + scratch[5].i = scratch[0].i + S_MUL(scratch[7].i,ya.r) + S_MUL(scratch[8].i,yb.r); + + scratch[6].r = S_MUL(scratch[10].i,ya.i) + S_MUL(scratch[9].i,yb.i); + scratch[6].i = -S_MUL(scratch[10].r,ya.i) - S_MUL(scratch[9].r,yb.i); + + C_SUB(*Fout1,scratch[5],scratch[6]); + C_ADD(*Fout4,scratch[5],scratch[6]); + + scratch[11].r = scratch[0].r + S_MUL(scratch[7].r,yb.r) + S_MUL(scratch[8].r,ya.r); + scratch[11].i = scratch[0].i + S_MUL(scratch[7].i,yb.r) + S_MUL(scratch[8].i,ya.r); + scratch[12].r = - S_MUL(scratch[10].i,yb.i) + S_MUL(scratch[9].i,ya.i); + scratch[12].i = S_MUL(scratch[10].r,yb.i) - S_MUL(scratch[9].r,ya.i); + + C_ADD(*Fout2,scratch[11],scratch[12]); + C_SUB(*Fout3,scratch[11],scratch[12]); + + ++Fout0;++Fout1;++Fout2;++Fout3;++Fout4; + } +} + +/* perform the butterfly for one stage of a mixed radix FFT */ +static void kf_bfly_generic( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + int m, + int p + ) +{ + int u,k,q1,q; + kiss_fft_cpx * twiddles = st->twiddles; + kiss_fft_cpx t; + kiss_fft_cpx scratchbuf[17]; + int Norig = st->nfft; + + /*CHECKBUF(scratchbuf,nscratchbuf,p);*/ + if (p>17) + celt_fatal("KissFFT: max radix supported is 17"); + + for ( u=0; u<m; ++u ) { + k=u; + for ( q1=0 ; q1<p ; ++q1 ) { + scratchbuf[q1] = Fout[ k ]; + if (!st->inverse) { + C_FIXDIV(scratchbuf[q1],p); + } + k += m; + } + + k=u; + for ( q1=0 ; q1<p ; ++q1 ) { + int twidx=0; + Fout[ k ] = scratchbuf[0]; + for (q=1;q<p;++q ) { + twidx += fstride * k; + if (twidx>=Norig) twidx-=Norig; + C_MUL(t,scratchbuf[q] , twiddles[twidx] ); + C_ADDTO( Fout[ k ] ,t); + } + k += m; + } + } +} + +static +void kf_shuffle( + kiss_fft_cpx * Fout, + const kiss_fft_cpx * f, + const size_t fstride, + int in_stride, + int * factors, + const kiss_fft_cfg st + ) +{ + const int p=*factors++; /* the radix */ + const int m=*factors++; /* stage's fft length/p */ + + /*printf ("fft %d %d %d %d %d %d\n", p*m, m, p, s2, fstride*in_stride, N);*/ + if (m==1) + { + int j; + for (j=0;j<p;j++) + { + Fout[j] = *f; + f += fstride*in_stride; + } + } else { + int j; + for (j=0;j<p;j++) + { + kf_shuffle( Fout , f, fstride*p, in_stride, factors,st); + f += fstride*in_stride; + Fout += m; + } + } +} + +static +void kf_work( + kiss_fft_cpx * Fout, + const kiss_fft_cpx * f, + const size_t fstride, + int in_stride, + int * factors, + const kiss_fft_cfg st, + int N, + int s2, + int m2 + ) +{ + int i; + kiss_fft_cpx * Fout_beg=Fout; + const int p=*factors++; /* the radix */ + const int m=*factors++; /* stage's fft length/p */ +#if 0 + /*printf ("fft %d %d %d %d %d %d\n", p*m, m, p, s2, fstride*in_stride, N);*/ + if (m==1) + { + /* int j; + for (j=0;j<p;j++) + { + Fout[j] = *f; + f += fstride*in_stride; + }*/ + } else { + int j; + for (j=0;j<p;j++) + { + kf_work( Fout , f, fstride*p, in_stride, factors,st, N*p, fstride*in_stride, m); + f += fstride*in_stride; + Fout += m; + } + } + + Fout=Fout_beg; + + switch (p) { + case 2: kf_bfly2(Fout,fstride,st,m); break; + case 3: kf_bfly3(Fout,fstride,st,m); break; + case 4: kf_bfly4(Fout,fstride,st,m); break; + case 5: kf_bfly5(Fout,fstride,st,m); break; + default: kf_bfly_generic(Fout,fstride,st,m,p); break; + } +#else + /*printf ("fft %d %d %d %d %d %d %d\n", p*m, m, p, s2, fstride*in_stride, N, m2);*/ + if (m==1) + { + /*for (i=0;i<N;i++) + { + int j; + Fout = Fout_beg+i*m2; + const kiss_fft_cpx * f2 = f+i*s2; + for (j=0;j<p;j++) + { + *Fout++ = *f2; + f2 += fstride*in_stride; + } + }*/ + }else{ + kf_work( Fout , f, fstride*p, in_stride, factors,st, N*p, fstride*in_stride, m); + } + + + + + switch (p) { + case 2: kf_bfly2(Fout,fstride,st,m, N, m2); break; + case 3: for (i=0;i<N;i++){Fout=Fout_beg+i*m2; kf_bfly3(Fout,fstride,st,m);} break; + case 4: kf_bfly4(Fout,fstride,st,m, N, m2); break; + case 5: for (i=0;i<N;i++){Fout=Fout_beg+i*m2; kf_bfly5(Fout,fstride,st,m);} break; + default: for (i=0;i<N;i++){Fout=Fout_beg+i*m2; kf_bfly_generic(Fout,fstride,st,m,p);} break; + } +#endif +} + +/* facbuf is populated by p1,m1,p2,m2, ... + where + p[i] * m[i] = m[i-1] + m0 = n */ +static +void kf_factor(int n,int * facbuf) +{ + int p=4; + + /*factor out powers of 4, powers of 2, then any remaining primes */ + do { + while (n % p) { + switch (p) { + case 4: p = 2; break; + case 2: p = 3; break; + default: p += 2; break; + } + if (p>32000 || (celt_int32_t)p*(celt_int32_t)p > n) + p = n; /* no more factors, skip to end */ + } + n /= p; + *facbuf++ = p; + *facbuf++ = n; + } while (n > 1); +} +/* + * + * User-callable function to allocate all necessary storage space for the fft. + * + * The return value is a contiguous block of memory, allocated with malloc. As such, + * It can be freed with free(), rather than a kiss_fft-specific function. + * */ +kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem ) +{ + kiss_fft_cfg st=NULL; + size_t memneeded = sizeof(struct kiss_fft_state) + + sizeof(kiss_fft_cpx)*(nfft-1); /* twiddle factors*/ + + if ( lenmem==NULL ) { + st = ( kiss_fft_cfg)KISS_FFT_MALLOC( memneeded ); + }else{ + if (mem != NULL && *lenmem >= memneeded) + st = (kiss_fft_cfg)mem; + *lenmem = memneeded; + } + if (st) { + int i; + st->nfft=nfft; + st->inverse = inverse_fft; +#ifdef FIXED_POINT + for (i=0;i<nfft;++i) { + celt_word32_t phase = i; + if (!st->inverse) + phase = -phase; + kf_cexp2(st->twiddles+i, DIV32(SHL32(phase,17),nfft)); + } +#else + for (i=0;i<nfft;++i) { + const double pi=3.14159265358979323846264338327; + double phase = ( -2*pi /nfft ) * i; + if (st->inverse) + phase *= -1; + kf_cexp(st->twiddles+i, phase ); + } +#endif + kf_factor(nfft,st->factors); + } + return st; +} + + + + +void kiss_fft_stride(kiss_fft_cfg st,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int in_stride) +{ + if (fin == fout) + { + celt_fatal("In-place FFT not supported"); + /*CHECKBUF(tmpbuf,ntmpbuf,st->nfft); + kf_work(tmpbuf,fin,1,in_stride, st->factors,st); + SPEEX_MOVE(fout,tmpbuf,st->nfft);*/ + } else { + kf_shuffle( fout, fin, 1,in_stride, st->factors,st); + kf_work( fout, fin, 1,in_stride, st->factors,st, 1, in_stride, 1); + } +} + +void kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout) +{ + kiss_fft_stride(cfg,fin,fout,1); +} + diff --git a/libcelt/kiss_fft.h b/libcelt/kiss_fft.h new file mode 100644 index 0000000..b5d6453 --- /dev/null +++ b/libcelt/kiss_fft.h @@ -0,0 +1,108 @@ +#ifndef KISS_FFT_H +#define KISS_FFT_H + +#include <stdlib.h> +#include <math.h> +#include "arch.h" + +#ifdef __cplusplus +extern "C" { +#endif + +/* + ATTENTION! + If you would like a : + -- a utility that will handle the caching of fft objects + -- real-only (no imaginary time component ) FFT + -- a multi-dimensional FFT + -- a command-line utility to perform ffts + -- a command-line utility to perform fast-convolution filtering + + Then see kfc.h kiss_fftr.h kiss_fftnd.h fftutil.c kiss_fastfir.c + in the tools/ directory. +*/ + +#ifdef USE_SIMD +# include <xmmintrin.h> +# define kiss_fft_scalar __m128 +#define KISS_FFT_MALLOC(nbytes) memalign(16,nbytes) +#else +#define KISS_FFT_MALLOC celt_alloc +#endif + + +#ifdef FIXED_POINT +#include "arch.h" +# define kiss_fft_scalar spx_int16_t +#else +# ifndef kiss_fft_scalar +/* default is float */ +# define kiss_fft_scalar float +# endif +#endif + +typedef struct { + kiss_fft_scalar r; + kiss_fft_scalar i; +}kiss_fft_cpx; + +typedef struct kiss_fft_state* kiss_fft_cfg; + +/* + * kiss_fft_alloc + * + * Initialize a FFT (or IFFT) algorithm's cfg/state buffer. + * + * typical usage: kiss_fft_cfg mycfg=kiss_fft_alloc(1024,0,NULL,NULL); + * + * The return value from fft_alloc is a cfg buffer used internally + * by the fft routine or NULL. + * + * If lenmem is NULL, then kiss_fft_alloc will allocate a cfg buffer using malloc. + * The returned value should be free()d when done to avoid memory leaks. + * + * The state can be placed in a user supplied buffer 'mem': + * If lenmem is not NULL and mem is not NULL and *lenmem is large enough, + * then the function places the cfg in mem and the size used in *lenmem + * and returns mem. + * + * If lenmem is not NULL and ( mem is NULL or *lenmem is not large enough), + * then the function returns NULL and places the minimum cfg + * buffer size in *lenmem. + * */ + +kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem); + +/* + * kiss_fft(cfg,in_out_buf) + * + * Perform an FFT on a complex input buffer. + * for a forward FFT, + * fin should be f[0] , f[1] , ... ,f[nfft-1] + * fout will be F[0] , F[1] , ... ,F[nfft-1] + * Note that each element is complex and can be accessed like + f[k].r and f[k].i + * */ +void kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout); + +/* + A more generic version of the above function. It reads its input from every Nth sample. + * */ +void kiss_fft_stride(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int fin_stride); + +/* If kiss_fft_alloc allocated a buffer, it is one contiguous + buffer and can be simply free()d when no longer needed*/ +#define kiss_fft_free celt_free + +/* + Cleans up some memory that gets managed internally. Not necessary to call, but it might clean up + your compiler output to call this before you exit. +*/ +void kiss_fft_cleanup(void); + + +#ifdef __cplusplus +} +#endif + +#endif diff --git a/libcelt/mdct.c b/libcelt/mdct.c index 12e9a0f..40ed07c 100644 --- a/libcelt/mdct.c +++ b/libcelt/mdct.c @@ -1,566 +1,160 @@ -/******************************************************************** - * * - * THIS FILE IS PART OF THE OggVorbis SOFTWARE CODEC SOURCE CODE. * - * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS * - * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE * - * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. * - * * - * THE OggVorbis SOURCE CODE IS (C) COPYRIGHT 1994-2002 * - * by the XIPHOPHORUS Company http://www.xiph.org/ * - * * - ******************************************************************** +/* (C) 2008 Jean-Marc Valin, CSIRO +*/ +/* + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions + are met: + + - Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + + - Redistributions in binary form must reproduce the above copyright + notice, this list of conditions and the following disclaimer in the + documentation and/or other materials provided with the distribution. + + - Neither the name of the Xiph.org Foundation nor the names of its + contributors may be used to endorse or promote products derived from + this software without specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR + CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, + EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, + PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR + PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF + LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING + NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +/* This is a simple MDCT implementation that uses a N/4 complex FFT + to do most of the work. It should be relatively straightforward to + plug in pretty much and FFT here. + + This replaces the Vorbis FFT (and uses the exact same API), which + was a bit too messy and that was ending up duplicating code + (might as well use the same FFT everywhere). + + The algorithm is similar to (and inspired from) Fabrice Bellard's + MDCT implementation in FFMPEG, but has differences in signs, ordering + and scaling in many places. +*/ - function: normalized modified discrete cosine transform - power of two length transform only [64 <= n ] - last mod: $Id: mdct.c 7187 2004-07-20 07:24:27Z xiphmont $ - - Original algorithm adapted long ago from _The use of multirate filter - banks for coding of high quality digital audio_, by T. Sporer, - K. Brandenburg and B. Edler, collection of the European Signal - Processing Conference (EUSIPCO), Amsterdam, June 1992, Vol.1, pp - 211-214 - - The below code implements an algorithm that no longer looks much like - that presented in the paper, but the basic structure remains if you - dig deep enough to see it. - - This module DOES NOT INCLUDE code to generate/apply the window - function. Everybody has their own weird favorite including me... I - happen to like the properties of y=sin(.5PI*sin^2(x)), but others may - vehemently disagree. - - ********************************************************************/ - -/* this can also be run as an integer transform by uncommenting a - define in mdct.h; the integerization is a first pass and although - it's likely stable for Vorbis, the dynamic range is constrained and - roundoff isn't done (so it's noisy). Consider it functional, but - only a starting point. There's no point on a machine with an FPU */ - -#include <stdio.h> -#include <stdlib.h> -#include <string.h> -#include <math.h> -/*#include "vorbis/codec.h"*/ #include "mdct.h" +#include "kiss_fft.h" +#include <math.h> #include "os_support.h" -/*#include "misc.h"*/ - -#define STIN static inline - -/* build lookups for trig functions; also pre-figure scaling and - some window function algebra. */ - -void mdct_init(mdct_lookup *lookup,int n){ - int *bitrev=celt_alloc(sizeof(*bitrev)*(n/4)); - DATA_TYPE *T=celt_alloc(sizeof(*T)*(n+n/4)); - - int i; - int n2=n>>1; - int log2n=lookup->log2n=rint(log((float)n)/log(2.f)); - lookup->n=n; - lookup->trig=T; - lookup->bitrev=bitrev; - -/* trig lookups... */ - - for(i=0;i<n/4;i++){ - T[i*2]=FLOAT_CONV(cos((M_PI/n)*(4*i))); - T[i*2+1]=FLOAT_CONV(-sin((M_PI/n)*(4*i))); - T[n2+i*2]=FLOAT_CONV(cos((M_PI/(2*n))*(2*i+1))); - T[n2+i*2+1]=FLOAT_CONV(sin((M_PI/(2*n))*(2*i+1))); - } - for(i=0;i<n/8;i++){ - T[n+i*2]=FLOAT_CONV(cos((M_PI/n)*(4*i+2))*.5); - T[n+i*2+1]=FLOAT_CONV(-sin((M_PI/n)*(4*i+2))*.5); - } - - /* bitreverse lookup... */ - - { - int mask=(1<<(log2n-1))-1,i,j; - int msb=1<<(log2n-2); - for(i=0;i<n/8;i++){ - int acc=0; - for(j=0;msb>>j;j++) - if((msb>>j)&i)acc|=1<<j; - bitrev[i*2]=((~acc)&mask)-1; - bitrev[i*2+1]=acc; - - } - } - lookup->scale=FLOAT_CONV(4.f/n); -} - -/* 8 point butterfly (in place, 4 register) */ -STIN void mdct_butterfly_8(DATA_TYPE *x){ - REG_TYPE r0 = x[6] + x[2]; - REG_TYPE r1 = x[6] - x[2]; - REG_TYPE r2 = x[4] + x[0]; - REG_TYPE r3 = x[4] - x[0]; - - x[6] = r0 + r2; - x[4] = r0 - r2; - - r0 = x[5] - x[1]; - r2 = x[7] - x[3]; - x[0] = r1 + r0; - x[2] = r1 - r0; - - r0 = x[5] + x[1]; - r1 = x[7] + x[3]; - x[3] = r2 + r3; - x[1] = r2 - r3; - x[7] = r1 + r0; - x[5] = r1 - r0; - -} - -/* 16 point butterfly (in place, 4 register) */ -STIN void mdct_butterfly_16(DATA_TYPE *x){ - REG_TYPE r0 = x[1] - x[9]; - REG_TYPE r1 = x[0] - x[8]; - - x[8] += x[0]; - x[9] += x[1]; - x[0] = MULT_NORM((r0 + r1) * cPI2_8); - x[1] = MULT_NORM((r0 - r1) * cPI2_8); - - r0 = x[3] - x[11]; - r1 = x[10] - x[2]; - x[10] += x[2]; - x[11] += x[3]; - x[2] = r0; - x[3] = r1; - - r0 = x[12] - x[4]; - r1 = x[13] - x[5]; - x[12] += x[4]; - x[13] += x[5]; - x[4] = MULT_NORM((r0 - r1) * cPI2_8); - x[5] = MULT_NORM((r0 + r1) * cPI2_8); - - r0 = x[14] - x[6]; - r1 = x[15] - x[7]; - x[14] += x[6]; - x[15] += x[7]; - x[6] = r0; - x[7] = r1; - - mdct_butterfly_8(x); - mdct_butterfly_8(x+8); -} - -/* 32 point butterfly (in place, 4 register) */ -STIN void mdct_butterfly_32(DATA_TYPE *x){ - REG_TYPE r0 = x[30] - x[14]; - REG_TYPE r1 = x[31] - x[15]; - - x[30] += x[14]; - x[31] += x[15]; - x[14] = r0; - x[15] = r1; - - r0 = x[28] - x[12]; - r1 = x[29] - x[13]; - x[28] += x[12]; - x[29] += x[13]; - x[12] = MULT_NORM( r0 * cPI1_8 - r1 * cPI3_8 ); - x[13] = MULT_NORM( r0 * cPI3_8 + r1 * cPI1_8 ); - - r0 = x[26] - x[10]; - r1 = x[27] - x[11]; - x[26] += x[10]; - x[27] += x[11]; - x[10] = MULT_NORM(( r0 - r1 ) * cPI2_8); - x[11] = MULT_NORM(( r0 + r1 ) * cPI2_8); - - r0 = x[24] - x[8]; - r1 = x[25] - x[9]; - x[24] += x[8]; - x[25] += x[9]; - x[8] = MULT_NORM( r0 * cPI3_8 - r1 * cPI1_8 ); - x[9] = MULT_NORM( r1 * cPI3_8 + r0 * cPI1_8 ); - - r0 = x[22] - x[6]; - r1 = x[7] - x[23]; - x[22] += x[6]; - x[23] += x[7]; - x[6] = r1; - x[7] = r0; - - r0 = x[4] - x[20]; - r1 = x[5] - x[21]; - x[20] += x[4]; - x[21] += x[5]; - x[4] = MULT_NORM( r1 * cPI1_8 + r0 * cPI3_8 ); - x[5] = MULT_NORM( r1 * cPI3_8 - r0 * cPI1_8 ); - - r0 = x[2] - x[18]; - r1 = x[3] - x[19]; - x[18] += x[2]; - x[19] += x[3]; - x[2] = MULT_NORM(( r1 + r0 ) * cPI2_8); - x[3] = MULT_NORM(( r1 - r0 ) * cPI2_8); - - r0 = x[0] - x[16]; - r1 = x[1] - x[17]; - x[16] += x[0]; - x[17] += x[1]; - x[0] = MULT_NORM( r1 * cPI3_8 + r0 * cPI1_8 ); - x[1] = MULT_NORM( r1 * cPI1_8 - r0 * cPI3_8 ); - - mdct_butterfly_16(x); - mdct_butterfly_16(x+16); +void mdct_init(mdct_lookup *l,int N) +{ + int i; + int N2, N4; + l->n = N; + N2 = N/2; + N4 = N/4; + l->kfft = kiss_fft_alloc(N4, 0, NULL, NULL); + l->ikfft = kiss_fft_alloc(N4, 1, NULL, NULL); + l->trig = celt_alloc(N2*sizeof(float)); + /* We have enough points that sine isn't necessary */ + for (i=0;i<N2;i++) + l->trig[i] = cos(2*M_PI*(i+1./8.)/N); + l->scale = 1./N4; } -/* N point first stage butterfly (in place, 2 register) */ -STIN void mdct_butterfly_first(DATA_TYPE *T, - DATA_TYPE *x, - int points){ - - DATA_TYPE *x1 = x + points - 8; - DATA_TYPE *x2 = x + (points>>1) - 8; - REG_TYPE r0; - REG_TYPE r1; - - do{ - - r0 = x1[6] - x2[6]; - r1 = x1[7] - x2[7]; - x1[6] += x2[6]; - x1[7] += x2[7]; - x2[6] = MULT_NORM(r1 * T[1] + r0 * T[0]); - x2[7] = MULT_NORM(r1 * T[0] - r0 * T[1]); - - r0 = x1[4] - x2[4]; - r1 = x1[5] - x2[5]; - x1[4] += x2[4]; - x1[5] += x2[5]; - x2[4] = MULT_NORM(r1 * T[5] + r0 * T[4]); - x2[5] = MULT_NORM(r1 * T[4] - r0 * T[5]); - - r0 = x1[2] - x2[2]; - r1 = x1[3] - x2[3]; - x1[2] += x2[2]; - x1[3] += x2[3]; - x2[2] = MULT_NORM(r1 * T[9] + r0 * T[8]); - x2[3] = MULT_NORM(r1 * T[8] - r0 * T[9]); - - r0 = x1[0] - x2[0]; - r1 = x1[1] - x2[1]; - x1[0] += x2[0]; - x1[1] += x2[1]; - x2[0] = MULT_NORM(r1 * T[13] + r0 * T[12]); - x2[1] = MULT_NORM(r1 * T[12] - r0 * T[13]); - - x1-=8; - x2-=8; - T+=16; - - }while(x2>=x); +void mdct_clear(mdct_lookup *l) +{ + kiss_fft_free(l->kfft); + kiss_fft_free(l->ikfft); + celt_free(l->trig); } -/* N/stage point generic N stage butterfly (in place, 2 register) */ -STIN void mdct_butterfly_generic(DATA_TYPE *T, - DATA_TYPE *x, - int points, - int trigint){ - - DATA_TYPE *x1 = x + points - 8; - DATA_TYPE *x2 = x + (points>>1) - 8; - REG_TYPE r0; - REG_TYPE r1; - - do{ - - r0 = x1[6] - x2[6]; - r1 = x1[7] - x2[7]; - x1[6] += x2[6]; - x1[7] += x2[7]; - x2[6] = MULT_NORM(r1 * T[1] + r0 * T[0]); - x2[7] = MULT_NORM(r1 * T[0] - r0 * T[1]); - - T+=trigint; - - r0 = x1[4] - x2[4]; - r1 = x1[5] - x2[5]; - x1[4] += x2[4]; - x1[5] += x2[5]; - x2[4] = MULT_NORM(r1 * T[1] + r0 * T[0]); - x2[5] = MULT_NORM(r1 * T[0] - r0 * T[1]); - - T+=trigint; - - r0 = x1[2] - x2[2]; - r1 = x1[3] - x2[3]; - x1[2] += x2[2]; - x1[3] += x2[3]; - x2[2] = MULT_NORM(r1 * T[1] + r0 * T[0]); - x2[3] = MULT_NORM(r1 * T[0] - r0 * T[1]); - - T+=trigint; - - r0 = x1[0] - x2[0]; - r1 = x1[1] - x2[1]; - x1[0] += x2[0]; - x1[1] += x2[1]; - x2[0] = MULT_NORM(r1 * T[1] + r0 * T[0]); - x2[1] = MULT_NORM(r1 * T[0] - r0 * T[1]); - - T+=trigint; - x1-=8; - x2-=8; - - }while(x2>=x); +void mdct_forward(mdct_lookup *l, float *in, float *out) +{ + int i; + int N, N2, N4, N8; + N = l->n; + N2 = N/2; + N4 = N/4; + N8 = N/8; + float f[N2]; + + /* Consider the input to be compused of four blocks: [a, b, c, d] */ + /* Shuffle, fold, pre-rotate (part 1) */ + for(i=0;i<N8;i++) + { + float re, im; + /* Real part arranged as -d-cR, Imag part arranged as -b+aR*/ + re = -in[N2+N4+2*i] - in[N2+N4-2*i-1]; + im = -in[N4+2*i] + in[N4-2*i-1]; + out[2*i] = re*l->trig[i] - im*l->trig[i+N4]; + out[2*i+1] = im*l->trig[i] + re*l->trig[i+N4]; + } + for(;i<N4;i++) + { + float re, im; + /* Real part arranged as a-bR, Imag part arranged as -c-dR */ + re = in[2*i-N4] - in[N2+N4-2*i-1]; + im = -(in[N4+2*i] + in[N+N4-2*i-1]); + out[2*i] = re*l->trig[i] - im*l->trig[i+N4]; + out[2*i+1] = im*l->trig[i] + re*l->trig[i+N4]; + } + + /* N/4 complex FFT, which should normally down-scale by 4/N (but doesn't now) */ + kiss_fft(l->kfft, (const kiss_fft_cpx *)out, (kiss_fft_cpx *)f); + + /* Post-rotate and apply the scaling if the FFT doesn't to it itself */ + for(i=0;i<N4;i++) + { + out[2*i] = l->scale * (-f[2*i+1]*l->trig[i+N4] + f[2*i] *l->trig[i]); + out[N2-1-2*i] = l->scale * (-f[2*i] *l->trig[i+N4] - f[2*i+1]*l->trig[i]); + } } -STIN void mdct_butterflies(mdct_lookup *init, - DATA_TYPE *x, - int points){ - - DATA_TYPE *T=init->trig; - int stages=init->log2n-5; - int i,j; - - if(--stages>0){ - mdct_butterfly_first(T,x,points); - } - - for(i=1;--stages>0;i++){ - for(j=0;j<(1<<i);j++) - mdct_butterfly_generic(T,x+(points>>i)*j,points>>i,4<<i); - } - - for(j=0;j<points;j+=32) - mdct_butterfly_32(x+j); +void mdct_backward(mdct_lookup *l, float *in, float *out) +{ + int i; + int N, N2, N4, N8; + N = l->n; + N2 = N/2; + N4 = N/4; + N8 = N/8; + float f[N2]; + + /* Pre-rotate */ + for(i=0;i<N4;i++) + { + out[2*i] = -in[N2-2*i-1] * l->trig[i] - in[2*i]*l->trig[i+N4]; + out[2*i+1] = in[N2-2*i-1] * l->trig[i+N4] - in[2*i]*l->trig[i]; + } + + /* Inverse N/4 complex FFT. This one should *not* downscale even in fixed-point */ + kiss_fft(l->ikfft, (const kiss_fft_cpx *)out, (kiss_fft_cpx *)f); + + /* Post-rotate */ + for(i=0;i<N4;i++) + { + float re, im; + re = f[2*i]; + im = f[2*i+1]; + f[2*i] = re*l->trig[i] + im*l->trig[i+N4]; + f[2*i+1] = im*l->trig[i] - re*l->trig[i+N4]; + } + /* De-shuffle the components for the middle of the window only */ + for(i = 0; i < N4; i++) + { + out[N4+2*i] =-f[2*i]; + out[N4+2*i+1] = f[N2-2*i-1]; + } + + /* Mirror on both sides for TDAC */ + for(i = 0; i < N4; i++) + { + out[i] =-out[N2-i-1]; + out[N-i-1] = out[N2+i]; + } } -void mdct_clear(mdct_lookup *l){ - if(l){ - if(l->trig)celt_free(l->trig); - if(l->bitrev)celt_free(l->bitrev); - memset(l,0,sizeof(*l)); - } -} - -STIN void mdct_bitreverse(mdct_lookup *init, - DATA_TYPE *x){ - int n = init->n; - int *bit = init->bitrev; - DATA_TYPE *w0 = x; - DATA_TYPE *w1 = x = w0+(n>>1); - DATA_TYPE *T = init->trig+n; - - do{ - DATA_TYPE *x0 = x+bit[0]; - DATA_TYPE *x1 = x+bit[1]; - - REG_TYPE r0 = x0[1] - x1[1]; - REG_TYPE r1 = x0[0] + x1[0]; - REG_TYPE r2 = MULT_NORM(r1 * T[0] + r0 * T[1]); - REG_TYPE r3 = MULT_NORM(r1 * T[1] - r0 * T[0]); - - w1 -= 4; - - r0 = HALVE(x0[1] + x1[1]); - r1 = HALVE(x0[0] - x1[0]); - - w0[0] = r0 + r2; - w1[2] = r0 - r2; - w0[1] = r1 + r3; - w1[3] = r3 - r1; - - x0 = x+bit[2]; - x1 = x+bit[3]; - - r0 = x0[1] - x1[1]; - r1 = x0[0] + x1[0]; - r2 = MULT_NORM(r1 * T[2] + r0 * T[3]); - r3 = MULT_NORM(r1 * T[3] - r0 * T[2]); - - r0 = HALVE(x0[1] + x1[1]); - r1 = HALVE(x0[0] - x1[0]); - - w0[2] = r0 + r2; - w1[0] = r0 - r2; - w0[3] = r1 + r3; - w1[1] = r3 - r1; - - T += 4; - bit += 4; - w0 += 4; - - }while(w0<w1); -} - -void mdct_backward(mdct_lookup *init, DATA_TYPE *in, DATA_TYPE *out){ - int n=init->n; - int n2=n>>1; - int n4=n>>2; - - /* rotate */ - - DATA_TYPE *iX = in+n2-7; - DATA_TYPE *oX = out+n2+n4; - DATA_TYPE *T = init->trig+n4; - - do{ - oX -= 4; - oX[0] = MULT_NORM(-iX[2] * T[3] - iX[0] * T[2]); - oX[1] = MULT_NORM (iX[0] * T[3] - iX[2] * T[2]); - oX[2] = MULT_NORM(-iX[6] * T[1] - iX[4] * T[0]); - oX[3] = MULT_NORM (iX[4] * T[1] - iX[6] * T[0]); - iX -= 8; - T += 4; - }while(iX>=in); - - iX = in+n2-8; - oX = out+n2+n4; - T = init->trig+n4; - - do{ - T -= 4; - oX[0] = MULT_NORM (iX[4] * T[3] + iX[6] * T[2]); - oX[1] = MULT_NORM (iX[4] * T[2] - iX[6] * T[3]); - oX[2] = MULT_NORM (iX[0] * T[1] + iX[2] * T[0]); - oX[3] = MULT_NORM (iX[0] * T[0] - iX[2] * T[1]); - iX -= 8; - oX += 4; - }while(iX>=in); - - mdct_butterflies(init,out+n2,n2); - mdct_bitreverse(init,out); - - /* roatate + window */ - - { - DATA_TYPE *oX1=out+n2+n4; - DATA_TYPE *oX2=out+n2+n4; - DATA_TYPE *iX =out; - T =init->trig+n2; - - do{ - oX1-=4; - - oX1[3] = MULT_NORM (iX[0] * T[1] - iX[1] * T[0]); - oX2[0] = -MULT_NORM (iX[0] * T[0] + iX[1] * T[1]); - - oX1[2] = MULT_NORM (iX[2] * T[3] - iX[3] * T[2]); - oX2[1] = -MULT_NORM (iX[2] * T[2] + iX[3] * T[3]); - - oX1[1] = MULT_NORM (iX[4] * T[5] - iX[5] * T[4]); - oX2[2] = -MULT_NORM (iX[4] * T[4] + iX[5] * T[5]); - - oX1[0] = MULT_NORM (iX[6] * T[7] - iX[7] * T[6]); - oX2[3] = -MULT_NORM (iX[6] * T[6] + iX[7] * T[7]); - - oX2+=4; - iX += 8; - T += 8; - }while(iX<oX1); - - iX=out+n2+n4; - oX1=out+n4; - oX2=oX1; - - do{ - oX1-=4; - iX-=4; - - oX2[0] = -(oX1[3] = iX[3]); - oX2[1] = -(oX1[2] = iX[2]); - oX2[2] = -(oX1[1] = iX[1]); - oX2[3] = -(oX1[0] = iX[0]); - - oX2+=4; - }while(oX2<iX); - - iX=out+n2+n4; - oX1=out+n2+n4; - oX2=out+n2; - do{ - oX1-=4; - oX1[0]= iX[3]; - oX1[1]= iX[2]; - oX1[2]= iX[1]; - oX1[3]= iX[0]; - iX+=4; - }while(oX1>oX2); - } -} - -void mdct_forward(mdct_lookup *init, DATA_TYPE *in, DATA_TYPE *out){ - int n=init->n; - int n2=n>>1; - int n4=n>>2; - int n8=n>>3; - DATA_TYPE *w=alloca(n*sizeof(*w)); /* forward needs working space */ - DATA_TYPE *w2=w+n2; - - /* rotate */ - - /* window + rotate + step 1 */ - - REG_TYPE r0; - REG_TYPE r1; - DATA_TYPE *x0=in+n2+n4; - DATA_TYPE *x1=x0+1; - DATA_TYPE *T=init->trig+n2; - - int i=0; - - for(i=0;i<n8;i+=2){ - x0 -=4; - T-=2; - r0= x0[2] + x1[0]; - r1= x0[0] + x1[2]; - w2[i]= MULT_NORM(r1*T[1] + r0*T[0]); - w2[i+1]= MULT_NORM(r1*T[0] - r0*T[1]); - x1 +=4; - } - - x1=in+1; - - for(;i<n2-n8;i+=2){ - T-=2; - x0 -=4; - r0= x0[2] - x1[0]; - r1= x0[0] - x1[2]; - w2[i]= MULT_NORM(r1*T[1] + r0*T[0]); - w2[i+1]= MULT_NORM(r1*T[0] - r0*T[1]); - x1 +=4; - } - - x0=in+n; - - for(;i<n2;i+=2){ - T-=2; - x0 -=4; - r0= -x0[2] - x1[0]; - r1= -x0[0] - x1[2]; - w2[i]= MULT_NORM(r1*T[1] + r0*T[0]); - w2[i+1]= MULT_NORM(r1*T[0] - r0*T[1]); - x1 +=4; - } - - - mdct_butterflies(init,w+n2,n2); - mdct_bitreverse(init,w); - - /* roatate + window */ - - T=init->trig+n2; - x0=out+n2; - - for(i=0;i<n4;i++){ - x0--; - out[i] =MULT_NORM((w[0]*T[0]+w[1]*T[1])*init->scale); - x0[0] =MULT_NORM((w[0]*T[1]-w[1]*T[0])*init->scale); - w+=2; - T+=2; - } -} diff --git a/libcelt/mdct.h b/libcelt/mdct.h index b91dc43..09762d4 100644 --- a/libcelt/mdct.h +++ b/libcelt/mdct.h @@ -1,83 +1,59 @@ -/******************************************************************** - * * - * THIS FILE IS PART OF THE OggVorbis SOFTWARE CODEC SOURCE CODE. * - * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS * - * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE * - * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. * - * * - * THE OggVorbis SOURCE CODE IS (C) COPYRIGHT 1994-2002 * - * by the XIPHOPHORUS Company http://www.xiph.org/ * - * * - ******************************************************************** - - function: modified discrete cosine transform prototypes - last mod: $Id: mdct.h 7187 2004-07-20 07:24:27Z xiphmont $ - - ********************************************************************/ - -#ifndef _OGG_mdct_H_ -#define _OGG_mdct_H_ - -/*#include "vorbis/codec.h"*/ - - - - - -/*#define MDCT_INTEGERIZED <- be warned there could be some hurt left here*/ -#ifdef MDCT_INTEGERIZED - -#define DATA_TYPE int -#define REG_TYPE register int -#define TRIGBITS 14 -#define cPI3_8 6270 -#define cPI2_8 11585 -#define cPI1_8 15137 - -#define FLOAT_CONV(x) ((int)((x)*(1<<TRIGBITS)+.5)) -#define MULT_NORM(x) ((x)>>TRIGBITS) -#define HALVE(x) ((x)>>1) - -#else - -#define DATA_TYPE float -#define REG_TYPE float -#define cPI3_8 .38268343236508977175F -#define cPI2_8 .70710678118654752441F -#define cPI1_8 .92387953251128675613F - -#define FLOAT_CONV(x) (x) -#define MULT_NORM(x) (x) -#define HALVE(x) ((x)*.5f) - -#endif - +/* (C) 2008 Jean-Marc Valin, CSIRO +*/ +/* + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions + are met: + + - Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + + - Redistributions in binary form must reproduce the above copyright + notice, this list of conditions and the following disclaimer in the + documentation and/or other materials provided with the distribution. + + - Neither the name of the Xiph.org Foundation nor the names of its + contributors may be used to endorse or promote products derived from + this software without specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR + CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, + EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, + PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR + PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF + LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING + NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +/* This is a simple MDCT implementation that uses a N/4 complex FFT + to do most of the work. It should be relatively straightforward to + plug in pretty much and FFT here. + + This replaces the Vorbis FFT (and uses the exact same API), which + was a bit too messy and that was ending up duplicating code + (might as well use the same FFT everywhere). + + The algorithm is similar to (and inspired from) Fabrice Bellard's + MDCT implementation in FFMPEG, but has differences in signs, ordering + and scaling in many places. +*/ + +#include "kiss_fft.h" typedef struct { - int n; - int log2n; - - DATA_TYPE *trig; - int *bitrev; - - DATA_TYPE scale; + int n; + kiss_fft_cfg kfft; + kiss_fft_cfg ikfft; + float *trig; + float scale; } mdct_lookup; -extern void mdct_init(mdct_lookup *lookup,int n); -extern void mdct_clear(mdct_lookup *l); -extern void mdct_forward(mdct_lookup *init, DATA_TYPE *in, DATA_TYPE *out); -extern void mdct_backward(mdct_lookup *init, DATA_TYPE *in, DATA_TYPE *out); - -#endif - - - - - - - - - - - +void mdct_init(mdct_lookup *l,int N); +void mdct_clear(mdct_lookup *l); +void mdct_forward(mdct_lookup *l, float *in, float *out); +void mdct_backward(mdct_lookup *l, float *in, float *out); diff --git a/libcelt/testcelt.c b/libcelt/testcelt.c index d552a9b..dcdcfbd 100644 --- a/libcelt/testcelt.c +++ b/libcelt/testcelt.c @@ -106,10 +106,15 @@ int main(int argc, char *argv[]) //printf ("\n"); //printf ("%d\n", len); /* This is to simulate packet loss */ +#if 1 if (rand()%100==-1) celt_decode(dec, NULL, len, out); else celt_decode(dec, data, len, out); +#else + for (i=0;i<frame_size*channels;i++) + out[i] = in[i]; +#endif //printf ("\n"); for (i=0;i<frame_size*channels;i++) rmsd += (in[i]-out[i])*1.0*(in[i]-out[i]); |