shithub: opus

Info  •  Files  •  Log  •  Branches

Inverse MDCT no longer requires any scratch space

--- a/celt/mdct.c

+++ b/celt/mdct.c

@@ -241,13 +241,10 @@

    int i;

    int N, N2, N4;

    kiss_twiddle_scalar sine;

-   VARDECL(kiss_fft_cpx, f2);

-   SAVE_STACK;

    N = l->n;

    N >>= shift;

    N2 = N>>1;

    N4 = N>>2;

-   ALLOC(f2, N4, kiss_fft_cpx);

    /* sin(x) ~= x here */

 #ifdef FIXED_POINT

    sine = TRIG_UPSCALE*(QCONST16(0.7853981f, 15)+N2)/N;

@@ -260,27 +257,27 @@

       /* 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_cpx * OPUS_RESTRICT yp = f2;

+      kiss_fft_scalar * OPUS_RESTRICT yp = out+(overlap>>1);

       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++)

       {

+         int rev;

          kiss_fft_scalar yr, yi;

-         kiss_fft_cpx yc;

+         rev = *bitrev++;

          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. We swap real and imag because we

             use an FFT instead of an IFFT. */

-         yc.i = yr - S_MUL(yi,sine);

-         yc.r = yi + S_MUL(yr,sine);

+         yp[2*rev+1] = yr - S_MUL(yi,sine);

+         yp[2*rev] = yi + S_MUL(yr,sine);

          /* Storing the pre-rotation directly in the bitrev order. */

-         yp[*bitrev++] = yc;

          xp1+=2*stride;

          xp2-=2*stride;

       }

    }

-   opus_fft_impl(l->kfft[shift], f2);

+   opus_fft_impl(l->kfft[shift], (kiss_fft_cpx*)(out+(overlap>>1)));

    /* Post-rotate and de-shuffle from both ends of the buffer at once to make

       it in-place. */

@@ -295,8 +292,8 @@

          kiss_fft_scalar re, im, yr, yi;

          kiss_twiddle_scalar t0, t1;

          /* We swap real and imag because we're using an FFT instead of an IFFT. */

-         re = f2[i].i;

-         im = f2[i].r;

+         re = yp0[1];

+         im = yp0[0];

          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 */

@@ -303,8 +300,8 @@

          yr = S_MUL(re,t0) - S_MUL(im,t1);

          yi = S_MUL(im,t0) + S_MUL(re,t1);

          /* We swap real and imag because we're using an FFT instead of an IFFT. */

-         re = f2[N4-i-1].i;

-         im = f2[N4-i-1].r;

+         re = yp1[1];

+         im = yp1[0];

          /* works because the cos is nearly one */

          yp0[0] = -(yr - S_MUL(yi,sine));

          yp1[1] = yi + S_MUL(yr,sine);

@@ -340,5 +337,4 @@

          wp2--;

       }

    }

-   RESTORE_STACK;

 }