Chris@69: /* Copyright (c) 2007-2008 CSIRO
Chris@69:    Copyright (c) 2007-2008 Xiph.Org Foundation
Chris@69:    Written by Jean-Marc Valin */
Chris@69: /*
Chris@69:    Redistribution and use in source and binary forms, with or without
Chris@69:    modification, are permitted provided that the following conditions
Chris@69:    are met:
Chris@69: 
Chris@69:    - Redistributions of source code must retain the above copyright
Chris@69:    notice, this list of conditions and the following disclaimer.
Chris@69: 
Chris@69:    - Redistributions in binary form must reproduce the above copyright
Chris@69:    notice, this list of conditions and the following disclaimer in the
Chris@69:    documentation and/or other materials provided with the distribution.
Chris@69: 
Chris@69:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
Chris@69:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
Chris@69:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
Chris@69:    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
Chris@69:    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
Chris@69:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
Chris@69:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
Chris@69:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
Chris@69:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
Chris@69:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
Chris@69:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
Chris@69: */
Chris@69: 
Chris@69: /* This is a simple MDCT implementation that uses a N/4 complex FFT
Chris@69:    to do most of the work. It should be relatively straightforward to
Chris@69:    plug in pretty much and FFT here.
Chris@69: 
Chris@69:    This replaces the Vorbis FFT (and uses the exact same API), which
Chris@69:    was a bit too messy and that was ending up duplicating code
Chris@69:    (might as well use the same FFT everywhere).
Chris@69: 
Chris@69:    The algorithm is similar to (and inspired from) Fabrice Bellard's
Chris@69:    MDCT implementation in FFMPEG, but has differences in signs, ordering
Chris@69:    and scaling in many places.
Chris@69: */
Chris@69: #ifndef __MDCT_MIPSR1_H__
Chris@69: #define __MDCT_MIPSR1_H__
Chris@69: 
Chris@69: #ifndef SKIP_CONFIG_H
Chris@69: #ifdef HAVE_CONFIG_H
Chris@69: #include "config.h"
Chris@69: #endif
Chris@69: #endif
Chris@69: 
Chris@69: #include "mdct.h"
Chris@69: #include "kiss_fft.h"
Chris@69: #include "_kiss_fft_guts.h"
Chris@69: #include <math.h>
Chris@69: #include "os_support.h"
Chris@69: #include "mathops.h"
Chris@69: #include "stack_alloc.h"
Chris@69: 
Chris@69: /* Forward MDCT trashes the input array */
Chris@69: #define OVERRIDE_clt_mdct_forward
Chris@69: void clt_mdct_forward(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scalar * OPUS_RESTRICT out,
Chris@69:       const opus_val16 *window, int overlap, int shift, int stride, int arch)
Chris@69: {
Chris@69:    int i;
Chris@69:    int N, N2, N4;
Chris@69:    VARDECL(kiss_fft_scalar, f);
Chris@69:    VARDECL(kiss_fft_cpx, f2);
Chris@69:    const kiss_fft_state *st = l->kfft[shift];
Chris@69:    const kiss_twiddle_scalar *trig;
Chris@69:    opus_val16 scale;
Chris@69: #ifdef FIXED_POINT
Chris@69:    /* Allows us to scale with MULT16_32_Q16(), which is faster than
Chris@69:       MULT16_32_Q15() on ARM. */
Chris@69:    int scale_shift = st->scale_shift-1;
Chris@69: #endif
Chris@69: 
Chris@69:     (void)arch;
Chris@69: 
Chris@69:    SAVE_STACK;
Chris@69:    scale = st->scale;
Chris@69: 
Chris@69:    N = l->n;
Chris@69:    trig = l->trig;
Chris@69:    for (i=0;i<shift;i++)
Chris@69:    {
Chris@69:       N >>= 1;
Chris@69:       trig += N;
Chris@69:    }
Chris@69:    N2 = N>>1;
Chris@69:    N4 = N>>2;
Chris@69: 
Chris@69:    ALLOC(f, N2, kiss_fft_scalar);
Chris@69:    ALLOC(f2, N4, kiss_fft_cpx);
Chris@69: 
Chris@69:    /* Consider the input to be composed of four blocks: [a, b, c, d] */
Chris@69:    /* Window, shuffle, fold */
Chris@69:    {
Chris@69:       /* Temp pointers to make it really clear to the compiler what we're doing */
Chris@69:       const kiss_fft_scalar * OPUS_RESTRICT xp1 = in+(overlap>>1);
Chris@69:       const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+N2-1+(overlap>>1);
Chris@69:       kiss_fft_scalar * OPUS_RESTRICT yp = f;
Chris@69:       const opus_val16 * OPUS_RESTRICT wp1 = window+(overlap>>1);
Chris@69:       const opus_val16 * OPUS_RESTRICT wp2 = window+(overlap>>1)-1;
Chris@69:       for(i=0;i<((overlap+3)>>2);i++)
Chris@69:       {
Chris@69:          /* Real part arranged as -d-cR, Imag part arranged as -b+aR*/
Chris@69:           *yp++ = S_MUL_ADD(*wp2, xp1[N2],*wp1,*xp2);
Chris@69:           *yp++ = S_MUL_SUB(*wp1, *xp1,*wp2, xp2[-N2]);
Chris@69:          xp1+=2;
Chris@69:          xp2-=2;
Chris@69:          wp1+=2;
Chris@69:          wp2-=2;
Chris@69:       }
Chris@69:       wp1 = window;
Chris@69:       wp2 = window+overlap-1;
Chris@69:       for(;i<N4-((overlap+3)>>2);i++)
Chris@69:       {
Chris@69:          /* Real part arranged as a-bR, Imag part arranged as -c-dR */
Chris@69:          *yp++ = *xp2;
Chris@69:          *yp++ = *xp1;
Chris@69:          xp1+=2;
Chris@69:          xp2-=2;
Chris@69:       }
Chris@69:       for(;i<N4;i++)
Chris@69:       {
Chris@69:          /* Real part arranged as a-bR, Imag part arranged as -c-dR */
Chris@69:           *yp++ =  S_MUL_SUB(*wp2, *xp2, *wp1, xp1[-N2]);
Chris@69:           *yp++ = S_MUL_ADD(*wp2, *xp1, *wp1, xp2[N2]);
Chris@69:          xp1+=2;
Chris@69:          xp2-=2;
Chris@69:          wp1+=2;
Chris@69:          wp2-=2;
Chris@69:       }
Chris@69:    }
Chris@69:    /* Pre-rotation */
Chris@69:    {
Chris@69:       kiss_fft_scalar * OPUS_RESTRICT yp = f;
Chris@69:       const kiss_twiddle_scalar *t = &trig[0];
Chris@69:       for(i=0;i<N4;i++)
Chris@69:       {
Chris@69:          kiss_fft_cpx yc;
Chris@69:          kiss_twiddle_scalar t0, t1;
Chris@69:          kiss_fft_scalar re, im, yr, yi;
Chris@69:          t0 = t[i];
Chris@69:          t1 = t[N4+i];
Chris@69:          re = *yp++;
Chris@69:          im = *yp++;
Chris@69: 
Chris@69:          yr = S_MUL_SUB(re,t0,im,t1);
Chris@69:          yi = S_MUL_ADD(im,t0,re,t1);
Chris@69: 
Chris@69:          yc.r = yr;
Chris@69:          yc.i = yi;
Chris@69:          yc.r = PSHR32(MULT16_32_Q16(scale, yc.r), scale_shift);
Chris@69:          yc.i = PSHR32(MULT16_32_Q16(scale, yc.i), scale_shift);
Chris@69:          f2[st->bitrev[i]] = yc;
Chris@69:       }
Chris@69:    }
Chris@69: 
Chris@69:    /* N/4 complex FFT, does not downscale anymore */
Chris@69:    opus_fft_impl(st, f2);
Chris@69: 
Chris@69:    /* Post-rotate */
Chris@69:    {
Chris@69:       /* Temp pointers to make it really clear to the compiler what we're doing */
Chris@69:       const kiss_fft_cpx * OPUS_RESTRICT fp = f2;
Chris@69:       kiss_fft_scalar * OPUS_RESTRICT yp1 = out;
Chris@69:       kiss_fft_scalar * OPUS_RESTRICT yp2 = out+stride*(N2-1);
Chris@69:       const kiss_twiddle_scalar *t = &trig[0];
Chris@69:       /* Temp pointers to make it really clear to the compiler what we're doing */
Chris@69:       for(i=0;i<N4;i++)
Chris@69:       {
Chris@69:          kiss_fft_scalar yr, yi;
Chris@69:          yr = S_MUL_SUB(fp->i,t[N4+i] , fp->r,t[i]);
Chris@69:          yi = S_MUL_ADD(fp->r,t[N4+i] ,fp->i,t[i]);
Chris@69:          *yp1 = yr;
Chris@69:          *yp2 = yi;
Chris@69:          fp++;
Chris@69:          yp1 += 2*stride;
Chris@69:          yp2 -= 2*stride;
Chris@69:       }
Chris@69:    }
Chris@69:    RESTORE_STACK;
Chris@69: }
Chris@69: 
Chris@69: #define OVERRIDE_clt_mdct_backward
Chris@69: void clt_mdct_backward(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scalar * OPUS_RESTRICT out,
Chris@69:       const opus_val16 * OPUS_RESTRICT window, int overlap, int shift, int stride, int arch)
Chris@69: {
Chris@69:    int i;
Chris@69:    int N, N2, N4;
Chris@69:    const kiss_twiddle_scalar *trig;
Chris@69: 
Chris@69:     (void)arch;
Chris@69: 
Chris@69:    N = l->n;
Chris@69:    trig = l->trig;
Chris@69:    for (i=0;i<shift;i++)
Chris@69:    {
Chris@69:       N >>= 1;
Chris@69:       trig += N;
Chris@69:    }
Chris@69:    N2 = N>>1;
Chris@69:    N4 = N>>2;
Chris@69: 
Chris@69:    /* Pre-rotate */
Chris@69:    {
Chris@69:       /* Temp pointers to make it really clear to the compiler what we're doing */
Chris@69:       const kiss_fft_scalar * OPUS_RESTRICT xp1 = in;
Chris@69:       const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+stride*(N2-1);
Chris@69:       kiss_fft_scalar * OPUS_RESTRICT yp = out+(overlap>>1);
Chris@69:       const kiss_twiddle_scalar * OPUS_RESTRICT t = &trig[0];
Chris@69:       const opus_int16 * OPUS_RESTRICT bitrev = l->kfft[shift]->bitrev;
Chris@69:       for(i=0;i<N4;i++)
Chris@69:       {
Chris@69:          int rev;
Chris@69:          kiss_fft_scalar yr, yi;
Chris@69:          rev = *bitrev++;
Chris@69:          yr = S_MUL_ADD(*xp2, t[i] , *xp1, t[N4+i]);
Chris@69:          yi = S_MUL_SUB(*xp1, t[i] , *xp2, t[N4+i]);
Chris@69:          /* We swap real and imag because we use an FFT instead of an IFFT. */
Chris@69:          yp[2*rev+1] = yr;
Chris@69:          yp[2*rev] = yi;
Chris@69:          /* Storing the pre-rotation directly in the bitrev order. */
Chris@69:          xp1+=2*stride;
Chris@69:          xp2-=2*stride;
Chris@69:       }
Chris@69:    }
Chris@69: 
Chris@69:    opus_fft_impl(l->kfft[shift], (kiss_fft_cpx*)(out+(overlap>>1)));
Chris@69: 
Chris@69:    /* Post-rotate and de-shuffle from both ends of the buffer at once to make
Chris@69:       it in-place. */
Chris@69:    {
Chris@69:       kiss_fft_scalar * OPUS_RESTRICT yp0 = out+(overlap>>1);
Chris@69:       kiss_fft_scalar * OPUS_RESTRICT yp1 = out+(overlap>>1)+N2-2;
Chris@69:       const kiss_twiddle_scalar *t = &trig[0];
Chris@69:       /* Loop to (N4+1)>>1 to handle odd N4. When N4 is odd, the
Chris@69:          middle pair will be computed twice. */
Chris@69:       for(i=0;i<(N4+1)>>1;i++)
Chris@69:       {
Chris@69:          kiss_fft_scalar re, im, yr, yi;
Chris@69:          kiss_twiddle_scalar t0, t1;
Chris@69:          /* We swap real and imag because we're using an FFT instead of an IFFT. */
Chris@69:          re = yp0[1];
Chris@69:          im = yp0[0];
Chris@69:          t0 = t[i];
Chris@69:          t1 = t[N4+i];
Chris@69:          /* We'd scale up by 2 here, but instead it's done when mixing the windows */
Chris@69:          yr = S_MUL_ADD(re,t0 , im,t1);
Chris@69:          yi = S_MUL_SUB(re,t1 , im,t0);
Chris@69:          /* We swap real and imag because we're using an FFT instead of an IFFT. */
Chris@69:          re = yp1[1];
Chris@69:          im = yp1[0];
Chris@69:          yp0[0] = yr;
Chris@69:          yp1[1] = yi;
Chris@69: 
Chris@69:          t0 = t[(N4-i-1)];
Chris@69:          t1 = t[(N2-i-1)];
Chris@69:          /* We'd scale up by 2 here, but instead it's done when mixing the windows */
Chris@69:          yr = S_MUL_ADD(re,t0,im,t1);
Chris@69:          yi = S_MUL_SUB(re,t1,im,t0);
Chris@69:          yp1[0] = yr;
Chris@69:          yp0[1] = yi;
Chris@69:          yp0 += 2;
Chris@69:          yp1 -= 2;
Chris@69:       }
Chris@69:    }
Chris@69: 
Chris@69:    /* Mirror on both sides for TDAC */
Chris@69:    {
Chris@69:       kiss_fft_scalar * OPUS_RESTRICT xp1 = out+overlap-1;
Chris@69:       kiss_fft_scalar * OPUS_RESTRICT yp1 = out;
Chris@69:       const opus_val16 * OPUS_RESTRICT wp1 = window;
Chris@69:       const opus_val16 * OPUS_RESTRICT wp2 = window+overlap-1;
Chris@69: 
Chris@69:       for(i = 0; i < overlap/2; i++)
Chris@69:       {
Chris@69:          kiss_fft_scalar x1, x2;
Chris@69:          x1 = *xp1;
Chris@69:          x2 = *yp1;
Chris@69:          *yp1++ = MULT16_32_Q15(*wp2, x2) - MULT16_32_Q15(*wp1, x1);
Chris@69:          *xp1-- = MULT16_32_Q15(*wp1, x2) + MULT16_32_Q15(*wp2, x1);
Chris@69:          wp1++;
Chris@69:          wp2--;
Chris@69:       }
Chris@69:    }
Chris@69: }
Chris@69: #endif /* __MDCT_MIPSR1_H__ */